{ "example": false, "notice": "Published catalog.", "items": [ { "capabilities": [ "avatar-video", "photo-to-video", "native-audio" ], "verified_on": "2026-10-16", "native_audio": true, "multi_shot": false, "t2v": false, "i2v": true, "open_source": false, "in_launch_feed": true, "name": "HeyGen Video", "summary": "HeyGen Video is the avatar engine family behind HeyGen, priced in credits on the enterprise page fetched 2026-10-16. It is a talking-presenter model, not a world text-to-video model like Veo or Kling.", "example": false, "price_band": "Enterprise 1 credit = $0.50; Avatar IV photo, twin, and studio 0.1 credits per second ($0.05 per second equivalent) on 2026-10-16", "where_to_run": [ "HeyGen enterprise API rates", "HeyGen developer platform", "HeyGen product app" ], "slug": "heygen-video", "markdown": "## Overview\n\nHeyGen Video, on this page, means the avatar video engines HeyGen bills as Avatar III, Avatar IV, and Avatar V, plus the separate Cinematic Avatar and Video Agent lines that sit on the same enterprise rate card. It does not mean a general text-to-video world model. If the brief is \"a street at night with no presenter,\" use [/models/veo](/models/veo), [/models/kling](/models/kling), or [/models/seedance](/models/seedance). If the brief is \"a person says this script on camera,\" you are in the right page. The product, the web plans, and the editor around those engines are [/tools/heygen](/tools/heygen). Plan dollars from the marketing site were not re-fetched for this model pass. They are not restated here.\n\nFilter tags are set from the enterprise page and the developer landing page opened on 2026-10-16. Native audio is true because the billed output is speaking avatar video. The rates are per second of that video, and the developer surface sells Avatar III, IV, and V as video engines rather than as silent animation. A separate \"generate a soundtrack\" flag was not a row. Speech is the product. `t2v` is false. A prompt-only world clip is not what Avatar IV meters. Video Agent is a different row, \"Prompt to Video\" at 0.0667 credits per second, and it is called out below so it is not silently folded into the avatar boolean. `i2v` is true only in the narrow sense the price card supports: Photo Avatar and the Avatar III \"photo-to-video pipeline\" turn an image into a talking video. That is not a free-form image-to-video model that will invent a new scene around a product still. `multi_shot` is false. The limits section allows up to 50 scenes in one composed video. A scene cap is an edit structure, not a Kling-style multi-shot prompt field. `open_source` is false.\n\nStatus: the enterprise rate card and the developer site are public. Self-serve versus enterprise is a commercial split the developer landing page names, without pasting the per-second table into the landing extract. This page treats the priced engines as generally available to customers who can open those accounts. A regional block list was not in the extracted enterprise text. A waitlist was not stated. Avatar III legacy v1/v2 is \"available to existing customers only and is not offered on the new developer platform.\" That legacy line is not GA for a new integration. The page says new work should use Avatar IV, Avatar V, or the current Avatar III engine.\n\n`www.heygen.com/developers` redirected to the developer marketing site. The saved landing lists Avatar III, IV, and V, Video Agent, translation, lipsync, and REST, CLI, and MCP. It does not include the per-second USD table. An older marketing figure of $0.10 per second was not on the pages saved in this pass. It is not quoted. Where the enterprise page itself prints an equivalence, that equivalence is used.\n\n## Specs\n\nVersion line. Avatar V is described as the highest-fidelity engine and supports Digital Twins only, at 0.1 credits per second. Avatar IV is the default v3 engine, with Photo Avatar, Digital Twin, and Studio Avatar each at 0.1 credits per second. Current Avatar III is a dedicated photo-to-video pipeline: Digital Twin and Studio Avatar at 0.0167 credits per second, Photo Avatar at 0.0433 credits per second. Legacy Avatar III v1/v2, existing customers only, is 0.0033 credits per second for photo, twin, and studio. Cinematic Avatar is not per second. It is a flat 7 credits per video, 4 to 15 seconds, 720p or 1080p only.\n\nEnterprise credit definition, printed on the page: 1 credit costs $0.50. The page then maps a self-serve comparison: Photo Avatar video at 0.1 credits per second is the equivalent of $0.05 per second. That $0.05 figure is HeyGen's arithmetic, quoted, not a second price list. Derived checks you can repeat: 0.0167 times $0.50 is $0.00835 per second, and 0.0433 times $0.50 is $0.02165 per second. Label those two derived. 7 credits times $0.50 is $3.50 per Cinematic Avatar video, derived. Video Agent at 0.0667 credits per second is about $0.03335 per second, derived.\n\nOutput limits extracted from the enterprise documentation:\n\n| Limit | Observed value |\n| --- | --- |\n| Frame rate | 25 fps for videos containing avatars. |\n| Resolution | Width and height each between 128 and 4,096 pixels. Default output 1080p. |\n| Aspect | 16:9 or 9:16. |\n| Scenes | Maximum 50 per video. |\n| Cinematic Avatar | 4 to 15 seconds, 720p or 1080p, flat 7 credits, not per second. |\n| Audio input | Maximum 10 minutes (600 seconds), in the limits section next to a 5,000 character text cap. This is an input cap, not the output-length cap. |\n| Output duration | Maximum 30 minutes per scene, on [usage limits](https://developers.heygen.com/docs/usage-limits), fetched 2026-10-16. The note says an avatar or image request to `POST /v3/videos` is one scene, so 30 minutes is that request's whole-video maximum. |\n| Video Agent prompt | 1 to 10,000 characters. Up to 20 file attachments. Types named: PNG, JPEG, MP4, WebM, MP3, WAV, and PDF. |\n\nThe 600-second line is an audio-input maximum. It is not the output-length cap. Output video specifications on the same usage-limits page say 30 minutes per scene. Speed 0.5× to 2.0×, with break tags, is labeled on that page as text-to-speech input for `POST /v3/voices/speech`, not as an Avatar IV field.\n\nHyperFrames rates on the same enterprise page are a composition meter, not an avatar model: 1080p at 30 fps is 0.1 credits per minute, 1080p at 60 fps is 0.2, 4K at 30 fps is 0.15, and 4K at 60 fps is 0.3. Video translation is billed per second of source, per target language. Dynamic duration can change the result. The rest of that translation sentence was not needed for the avatar spec. Translation is a feature, not Avatar IV.\n\nThe developer landing does not add resolution or duration numbers beyond the engine names. MCP and CLI are delivery methods for the same platform. They are not a second model.\n\n### Prompt notes\n\nAvatar jobs are scripts plus a portrait or a twin, not cinematic world prompts. The enterprise page does not include a sample script that this directory re-ran. Video Agent is the row that takes a long prompt and up to 20 attachments. Its rate is the 0.0667 credits per second line. If a teammate says \"HeyGen text-to-video,\" ask whether they mean Video Agent or an avatar reading a script. The filter on this page stays false for world text-to-video so Video Agent is not mistaken for Veo.\n\nNo arena score for Avatar IV or V was opened. Not observed. Talking-head quality and world-model arena scores are not comparable even when both exist.\n\n## Where to run\n\n1. Enterprise credit rates: [developers.heygen.com/docs/enterprise-pricing](https://developers.heygen.com/docs/enterprise-pricing). Use that page for credits per second. Output limits, including the 30-minute scene cap, are on [developers.heygen.com/docs/usage-limits](https://developers.heygen.com/docs/usage-limits).\n2. Developer marketing site: [developers.heygen.com](https://developers.heygen.com), which is where `https://www.heygen.com/developers` redirected. Engine names, Video Agent, translation, lipsync, and API styles. No per-second table in the extract.\n3. The product: [/tools/heygen](/tools/heygen), and the app behind it. Web plan prices were not re-opened on 2026-10-16.\n\nA non-HeyGen resale of Avatar IV, with a fal or Replicate price card, was not opened. Official only. The enterprise docs, the developer site, and the product are the three doors, all HeyGen. That satisfies the two-entry rule without inventing a second vendor.\n\n## Multi-channel pricing\n\nOne vendor, two commercial surfaces, and a gap where they were not both fully priced.\n\n**Enterprise API, fetched 2026-10-16.** 1 credit = $0.50. Avatar V digital twin 0.1 credits per second. Avatar IV photo, digital twin, and studio 0.1 credits per second, and the page states that 0.1 credits per second equals $0.05 per second on the self-serve comparison. Current Avatar III at 0.0167 or 0.0433 credits per second. Legacy v1/v2 at 0.0033 credits per second for existing customers. Cinematic Avatar 7 credits per video. Video Agent 0.0667 credits per second. HyperFrames per minute as listed in Specs. These are the verified numbers.\n\n**Developer landing, same day.** No dollar table. Do not fill the hole with a remembered $0.10 per second marketing rate. It was not on the saved landing.\n\n**Consumer web plans.** Not re-observed in this pass. If [/tools/heygen](/tools/heygen) lists Creator or other plan prices from a separate review, those figures stay on the tool page until the pricing URL is opened again. A plan credit and an enterprise credit can share the word \"credit\" and still differ. The enterprise page is explicit that you multiply its credit rate by $0.50 to compare with self-serve. Use that instruction rather than assuming every credit in the product is worth $0.50.\n\nWorked examples, derived except where the page already printed the equivalence. Ten seconds of Avatar IV at the printed $0.05 per second is $0.50. One minute is $3.00. Those products use HeyGen's own $0.05 equivalence. Ten seconds of Cinematic Avatar is still 7 credits, $3.50 derived, because the flat rate does not shrink for a shorter clip inside the 4 to 15 second window. A 4-second cinematic clip and a 15-second cinematic clip cost the same on this card. That is the main reason to read the row before you compare it with Avatar IV's per-second meter.\n\n## Workflow and platform fit\n\nLabels only.\n\n**Faceless.** A HeyGen avatar is the opposite of a faceless channel unless the face is a synthetic presenter you intend to show. For B-roll with no person, do not spend avatar seconds. Generate the street with another model and, if you need a narrator you never show, prefer [/tools/elevenlabs](/tools/elevenlabs) under stock or generated pictures. HeyGen fits faceless only when \"faceless\" means \"no human film crew,\" and a digital twin is acceptable on screen.\n\n**UGC ads.** This is a primary fit. A photo avatar or a digital twin can say an ad script in 9:16 at the default 1080p. [/tools/creatify](/tools/creatify) packages similar ad structures and may call HeyGen or its own pipeline. This page does not claim Creatify resells Avatar IV. That integration was not opened. Use HeyGen directly when you need the credit math above.\n\n**AI avatar.** This is the home workflow. Photo Avatar is the image-to-talking-video path. Digital Twin and Studio Avatar are the other two Avatar IV rates, at the same 0.1 credits per second. Avatar V narrows the input to digital twins. Pick V only when you already have a twin and you want the row the page calls highest fidelity. The page does not publish a side-by-side quality score.\n\n**Long-to-short.** Video translation and a 50-scene composition can repackage an existing piece. They do not detect highlights. Cut the long source in [/tools/capcut](/tools/capcut) or a clipper, then send the spoken segments you still need. HyperFrames is HeyGen's composition meter for programmatic pieces, billed per minute at the rates above, which is a different order of magnitude from avatar seconds. Read the unit. 0.1 credits per minute is not 0.1 credits per second.\n\n**Captions and cover.** Caption burn-in was not a field on the rate card. Assume captions are an editor step unless the product UI you are in shows otherwise. A cover is a still. The default video frame is 1080p, 16:9 or 9:16, which can be grabbed as a cover and is not itself a design template.\n\n**Voice and publish.** Audio is the reason you bought the second. Publish 9:16 for Shorts, Reels, and TikTok. 1:1 was not in the aspect list. 25 fps is the avatar frame rate in the limits, which is fine for those platforms and is not a 24 fps cinema master. Width and height must each sit between 128 and 4,096. Default 1080p is the sensible publish preset. Cinematic Avatar cannot be asked for 4K on this card. It is 720p or 1080p only.\n\n**Shorts vertical fit.** Strong. 9:16 is an official aspect, the default resolution is 1080p, and the meter is per second so a 15-second spoken Short is 15 times the engine rate. At the printed Avatar IV equivalence that is $0.75 derived for 15 seconds. You still add captions in an editor if the player chrome demands them. A silent visual gag with no speaker is the wrong engine.\n\n## Pros\n\n- The enterprise page prints both the credit rate and the $0.50 credit value, and it prints the $0.05 per second Photo Avatar equivalence itself.\n- 9:16 and 16:9 are the aspect pair a Shorts-plus-YouTube setup actually needs.\n- Default 1080p and a 128 to 4,096 pixel bound are concrete. You are not guessing the export size.\n- Avatar IV's three types share one rate, which makes photo versus studio a quality choice rather than a surprise invoice.\n- Cinematic Avatar's flat 7 credits is easy to compare with a long Avatar IV take. Short cinematic clips are expensive relative to their length. Long ones inside the 15-second cap are closer to the per-second engine.\n- Legacy v1/v2 is labeled existing-customers-only, which prevents a new integration from quoting the cheapest 0.0033 rate by accident.\n- Failure to confuse this model with world text-to-video is a feature of the rate card. The card is about avatars, agents, translation, and composition.\n\n## Cons\n\n- It will not make a general B-roll shot. Teams that brief HeyGen like Veo will burn credits on the wrong task or get a talking head they did not want.\n- Consumer plan dollars were not re-verified. An enterprise credit at $0.50 may not match a web-plan credit. The page tells you to multiply. It does not paste the current web grid.\n- A marketing $0.10 per second figure was not on the saved developer landing. Older posts that use it should not be mixed with the $0.05 equivalence without a new fetch.\n- The 10-minute figure is an audio-input cap. The output cap on the usage-limits page is 30 minutes per scene, which is a different number. Do not swap them.\n- Video Agent's prompt-to-video rate is easy to merge with Avatar IV. It is a different row and a different boolean.\n- HyperFrames is billed per minute. A spreadsheet that copies the avatar per-second rate onto a HyperFrames render will be wrong by a factor of sixty.\n- No second vendor's price card was opened. You cannot shop Avatar IV across fal and HeyGen from this research pass.\n- Multi-scene videos go to 50 scenes, and the card does not say whether every scene burns a full avatar rate. Scene-level billing detail is not observed.\n\n## Alternatives\n\n- [Kling](/models/kling) has an Avatar row on its own unit card, at $0.056 per second in 720p and $0.112 in 1080p, plus a general video model with native audio. Kling's avatar and HeyGen's Avatar IV are different products with different rigs. Compare a test, not just the second price.\n- [Veo](/models/veo), [Seedance](/models/seedance), and [Runway Gen](/models/runway-gen) make shots of the world. They do not replace a photo avatar with a script.\n- [Luma Ray](/models/luma-ray) can modify a performance you already filmed and keep its audio. It does not start from a still portrait and a script the way Photo Avatar does.\n- [MiniMax Hailuo](/models/minimax-hailuo) is a lower per-second world model without a documented presenter mode.\n\nCreatify is a tool alternative, not a model. It is linked below.\n\n## Related tools\n\n- [/tools/heygen](/tools/heygen) is the product. Web plans and the in-app path live there. This model page keeps the enterprise credit card.\n- [/tools/creatify](/tools/creatify) is the UGC-ad product when you want a template around a presenter rather than a raw avatar API call.\n- [/tools/elevenlabs](/tools/elevenlabs) is the voice swap or the unseen narrator when an avatar is the wrong picture.\n- [/tools/capcut](/tools/capcut) burns captions, builds covers, and cuts avatar clips into a Short.\n- [/tools/runway](/tools/runway), [/tools/kling](/tools/kling), [/tools/seedance](/tools/seedance), and [/tools/higgsfield](/tools/higgsfield) are the studios for the non-avatar shots in the same video.\n\n## FAQ\n\n### Is HeyGen a text-to-video model?\n\nNot in the world-model sense. Avatar engines turn a script plus a photo, twin, or studio avatar into a talking video. Video Agent is a separate \"Prompt to Video\" rate at 0.0667 credits per second. This page sets text-to-video to false so the avatar engines are not filtered in as Veo substitutes.\n\n### What does one second of Avatar IV cost?\n\n0.1 credits. The enterprise page says 1 credit costs $0.50 and that 0.1 credits per second equals $0.05 per second. Photo, digital twin, and studio Avatar IV share that rate. Fetched 2026-10-16.\n\n### Can I export 9:16?\n\nYes. The limits list aspect ratio 16:9 or 9:16, default output 1080p, 25 fps when the video contains avatars.\n\n### Is the cheap 0.0033 rate available to new users?\n\nThe page says legacy Avatar III v1/v2 is for existing customers only and is not offered on the new developer platform. New integrations should use Avatar IV, Avatar V, or the current Avatar III rates.\n\n### Were consumer subscription prices checked?\n\nNot in this pass. Use the enterprise card for API math. Re-open the web pricing page before you publish a monthly plan price. The tool page may hold an earlier note. It is a different review.", "path": "/models/heygen-video" }, { "capabilities": [ "text-to-video", "image-to-video", "video-to-video", "native-audio" ], "verified_on": "2026-10-16", "native_audio": true, "multi_shot": false, "t2v": true, "i2v": true, "open_source": false, "in_launch_feed": true, "name": "Higgsfield", "summary": "Higgsfield's own video models in the public API catalog on 2026-10-16 are Cinema Studio 4.0 and Genjutsu. Soul is the still family. DoP was not in that catalog snapshot. Dollar rates for these first-party models were not in the schemas.", "example": false, "price_band": "Official credit price for Cinema Studio 4.0 and Genjutsu not observed on 2026-10-16", "where_to_run": [ "Higgsfield API Cinema Studio 4.0", "Higgsfield API Genjutsu", "Higgsfield console" ], "slug": "higgsfield", "markdown": "## Overview\n\nThis page is about Higgsfield's own models. It is not about every model the Higgsfield studio can call. Seedance, Kling, Wan, and MiniMax routes on the same API are other companies' models. Their selection pages are [/models/seedance](/models/seedance), [/models/kling](/models/kling), and [/models/minimax-hailuo](/models/minimax-hailuo). The studio as a product, including plans and the MCP endpoint, is [/tools/higgsfield](/tools/higgsfield). Keeping those apart is the point of a model page.\n\nOn 2026-10-16 the public docs index `https://docs.higgsfield.ai/docs/llms-full.txt` was saved (about 596 KB, HTTP 200). Its video category copy names Genjutsu, Seedance, Kling, Wan, Happy Horse, Cinema Studio, MiniMax, LTX, PixVerse, and Grok. The first-party video pages inside that file are Cinema Studio 4.0 and Genjutsu. Soul, Soul Cinema, Soul ID, and SOUL V2 are image models. The word \"dop\" does not occur in that file. A direct request to `https://docs.higgsfield.ai/docs/guides/video` returned a 307 to `/docs/`. Search indexes still show an older guide that posts to `https://api.higgsfield.ai/higgsfield-ai/dop/standard` with a required `prompt` and `image_url`. That guide body was not saved as a 200 response in this environment. This page does not treat DoP as a current catalog model. It records the indexed endpoint as a historical lead that the live index did not confirm. Duration, resolution, and a dollar rate for DoP are not observed.\n\nStatus: Cinema Studio 4.0 and Genjutsu are documented public API workflows, with console playground links. That is general availability of the documented endpoints, subject to account access. The models index says the catalog documents 84 entries and that availability can change, and it points at `https://console.higgsfield.ai` for the account. A region block list was not in the extracted pages. A waitlist was not stated. `open_source` is false.\n\nFilter booleans describe Cinema Studio 4.0, the generative model, and they are the conservative reading of its schema. Text-to-video is true because the usage note says the workflow uses text-to-video when reference arrays are empty. Image-to-video is true because a non-empty image array switches the same endpoint to reference-to-video, with up to 30 image URLs. Native audio is true because the schema includes a boolean described as \"Generate audio with the video.\" Multi-shot is false. The pacing enum includes `single-shot` among `chaotic`, `dynamic`, `calm`, and `single-shot`. That is a style control, not a list of shot prompts. Camera move is a separate enum of named moves. Neither is a multi-shot API. Genjutsu does not change these booleans: it is video-to-video (motion transfer, object swap, restyle), so it is not a text-to-video world model, and the page's `t2v` bit follows Cinema Studio.\n\n## Specs\n\nVersion line. Cinema Studio 4.0 is the public generate endpoint, `POST https://api.higgsfield.ai/higgsfield/cinema-studio/4.0`, endpoint id `higgsfield/cinema-studio/4.0`. The usage notes say this endpoint produces a new video, and that editing and extension use the separate Seedance 2.5 endpoints. Read that sentence carefully. Cinema Studio is documented as its own workflow, and the same note sends edit and extend work to Seedance 2.5. This page does not claim Cinema Studio's weights are Seedance, and it does not hide the cross-reference. A \"runs on Seedance\" sentence from an indexed help article was not in the saved `llms-full` Cinema Studio chapter. That stronger claim is not observed in the saved body.\n\nCinema Studio 4.0 fields extracted from the saved schema:\n\n| Field | Observed value |\n| --- | --- |\n| Required | `prompt` (minimum 1 character, must contain non-whitespace). |\n| Mode | Text-to-video if media arrays are empty. Reference-to-video if any media array is non-empty. |\n| Duration | Integer seconds, minimum 4, maximum 30. |\n| Resolution | `480p` or `720p`. 1080p was not in the allowed values. |\n| Aspect | `16:9`, `4:3`, `1:1`, `3:4`, `9:16`, `21:9`. |\n| Audio output | Boolean, \"Generate audio with the video.\" |\n| Images | Up to 30 public URLs. |\n| Videos | Up to 10 public URLs. |\n| Audio inputs | Up to 10 public URLs. |\n| Combined references | At most 50 items. Video and audio may be normalized to separate 30-second total budgets. |\n| Prompt tokens | Optional one-based tokens such as image, video, and audio references. Each token must match an attached item of that type. |\n| Creative enums | Era, genre, light, pacing, lens look, camera body, a long scene-preset list, aperture, and a camera-move list. Omit a creative field to let the director choose. The literal string `auto` is not accepted for those enums. |\n\nFrame rate, a seed, and a dollar or credit cost were not in the extracted schema. Not observed.\n\nGenjutsu is the other first-party video family. Motion transfer is `POST https://api.higgsfield.ai/higgsfield/genjutsu/motion-transfer/v1.0`. The source video must be at least 4 seconds. Videos longer than 30 seconds are trimmed to 30. Output duration follows the prepared source. Provide 1 to 8 image references. The prompt is optional and defaults to empty. A legacy path spelled `higgsfiled` remains callable and is hidden from discovery. Object swap and restyle are sibling endpoints under `/higgsfield/genjutsu/`. Their full enums were not copied into this sheet. Genjutsu's price was not in the extract.\n\nSoul V2 is text-to-image, `POST` under the Soul standard path documented in the image guide. It is not a video model. Soul Cinema is another text-to-image endpoint. They matter on this page only so a reader does not brief Soul as a video generator. The image guide's example posts a landscape prompt and says completed images are retained at least seven days. Aspect and resolution for Soul were in an indexed image guide. The saved Soul chapter should be re-read before those still specs are quoted as video specs. They are not repeated here as Cinema Studio specs.\n\nBilling mechanics that do apply to every generation, from `https://docs.higgsfield.ai/docs/concepts/billing-and-retention.md`: charges hit successful requests in account credits. The exact cost depends on the model and the parameters. Credits expire one year after they are added. An estimate endpoint can be called before submit. The sample response on that page (`credits` 1.500 and `usd` 0.094) is explicitly an illustration of the response format for a Soul-shaped call, and the page says to use the estimate for your own account as the authoritative amount. It is not a Cinema Studio price. Requests that end `failed` or `nsfw` are not charged. A queued request can be canceled only before processing starts, and a successful cancel is refunded. Outputs remain available at least seven days.\n\nAuthentication, from the FAQ in the same docs set: base URL `https://api.higgsfield.ai`, header `Authorization: Key YOUR_KEY_ID:YOUR_KEY_SECRET`. Not a bearer token. Generation is asynchronous and returns `request_id`, `status_url`, and `cancel_url`.\n\n### Prompt notes\n\nCinema Studio's published quick start, not a clip rendered here, is \"A cinematic wide shot of ocean waves at sunset, with gentle camera movement.\" The schema then offers a camera-move enum that includes `dolly-in`, `drone-orbit`, `handheld`, `static-shot`, and many others. If you need a specific move, set the enum. The note says omitting the field lets the director choose, and sending the string `auto` is invalid. Reference tokens use a one-based syntax that must point at an attached image, video, or audio item. A token without the attachment is a request bug, not a style.\n\nDoP prompt behavior is not observed, because the DoP page was not in the saved catalog.\n\nArena scores for Cinema Studio or Genjutsu were not opened. Not observed.\n\n## Where to run\n\nOfficial only, two or more doors on Higgsfield's own platform:\n\n1. Cinema Studio 4.0 generate docs, saved inside [the docs index](https://docs.higgsfield.ai/docs/llms-full.txt) and addressed as `https://docs.higgsfield.ai/docs/models/cinema-studio-4/generate`. Endpoint `https://api.higgsfield.ai/higgsfield/cinema-studio/4.0`.\n2. Genjutsu motion transfer: `https://docs.higgsfield.ai/docs/models/genjutsu/motion-transfer` in the same index. Endpoint `https://api.higgsfield.ai/higgsfield/genjutsu/motion-transfer/v1.0`.\n3. The console the docs name as the place to discover models and open playgrounds: [console.higgsfield.ai](https://console.higgsfield.ai).\n4. The product page in this directory: [/tools/higgsfield](/tools/higgsfield).\n\nA non-Higgsfield host for Cinema Studio 4.0 or Genjutsu was not found on fal, Runway, or Luma in the files read for this pass. Resale is not observed. The two API workflows plus the console are the entry points, and they are one vendor. That matches the rule that a single-vendor model may say official only, as long as more than one door is linked. The console and the API are those doors.\n\nIndexed DoP: `POST https://api.higgsfield.ai/higgsfield-ai/dop/standard`. Not confirmed by the saved catalog. Do not ship a client against it without a current schema from the console.\n\n## Multi-channel pricing\n\nOfficial only. A dollar or credit figure for Cinema Studio 4.0 or for Genjutsu was not in the schemas or in the billing chapter. **Price: not observed.** Verified by absence in the files fetched on 2026-10-16, which is a weaker statement than a printed number and is the honest one.\n\nWhat you can say about money:\n\n- The billing chapter says the unit is account credits, model-specific, with a one-year expiry.\n- Failed and `nsfw` results are not charged. Canceled queued jobs are refunded.\n- The estimate API returns `credits` and `usd` for the call you submit. The printed Soul example is not transferable.\n- Consumer plan dollars on the Higgsfield marketing site were not re-fetched for this model pass. They may exist on the tool page from a separate review. They are not copied here as Cinema Studio rates.\n\nThere is no fal price, no Runway credit row, and no Luma credit row for Cinema Studio 4.0 in the extracts. Do not borrow Seedance's price just because the Cinema Studio note mentions Seedance 2.5 endpoints for edit and extend. Those endpoints have their own prices on the Seedance page, and they are a different request.\n\n## Workflow and platform fit\n\nLabels only. No unpublished workflow URLs.\n\n**Faceless.** Cinema Studio can run 4 to 30 seconds, 9:16, 480p or 720p, with a generate-audio flag. That is a complete faceless beat if the generated audio is usable. Usability was not tested. 720p is the top enum value, which is enough for a phone feed and short of a 1080p master. Captions are still [/tools/capcut](/tools/capcut). A named voice is still [/tools/elevenlabs](/tools/elevenlabs) when the boolean is not the right speaker.\n\n**UGC ads.** Reference-to-video with up to 30 images is the storyboard path. Genjutsu object swap is the \"replace the product in a clip you already have\" path, with a 4 to 30 second source window. [/tools/creatify](/tools/creatify) is the packaged UGC template if you do not want to drive these endpoints.\n\n**AI avatar.** Neither Cinema Studio nor Genjutsu is documented as a photo-avatar engine. Motion transfer puts a reference image's appearance onto source motion, which can resemble an avatar and is still a video-to-video job with a required source clip. For a presenter who speaks from a script, use [/models/heygen-video](/models/heygen-video).\n\n**Long-to-short.** Genjutsu trims sources over 30 seconds. It does not chapter a podcast. Cut the long file first, then restyle a short span.\n\n**Captions and cover.** Soul can make a still. Cinema Studio's response shape in the docs is a video URL. Grab a cover frame in CapCut. No caption field was in the schema.\n\n**Voice and publish.** Generate-audio is optional in the schema. Publish vertical by setting `9:16`. Inputs must be public URLs. The docs reject `asset://` on some sibling endpoints and tell you to upload files when you do not have a public URL. A private product shot must be uploaded through their file flow before reference-to-video will see it. Output retention is at least seven days. Download before you depend on the URL.\n\n**Shorts vertical fit.** This is one of the better documented vertical fits among first-party studio models: 9:16 is in the enum, duration covers a Shorts hook through a 30-second spot, and the top resolution is 720p rather than a surprise 4K bill. The missing price is the reason not to promise a unit cost. Run the estimate endpoint on the account that will pay.\n\n## Pros\n\n- The saved catalog separates first-party models from hosted Seedance and Kling, which is the distinction this directory also needs.\n- Cinema Studio's schema is specific: duration bounds, two resolutions, six aspects, and a generate-audio boolean, plus hard reference caps.\n- 9:16 is a first-class aspect, and 30 seconds is inside the duration maximum.\n- Genjutsu states a minimum and a trim point for source video, so a motion-transfer job has a known length rule.\n- Billing rules for failure, NSFW, cancel, and seven-day retention are written down, even though the model price is not.\n- The docs tell you to call the estimate endpoint instead of trusting a blog price. That is the correct instruction while the printed example is labeled illustrative.\n\n## Cons\n\n- The actual credit and dollar cost of Cinema Studio 4.0 and Genjutsu was not in the public schema. A selection page cannot finish a budget from the docs alone.\n- 1080p is not an allowed Cinema Studio resolution in the extract. A brand master at 1080p is a different model or an upscale. Not observed as a Cinema Studio tier.\n- DoP is in indexed guides and absent from the saved full index. An integration written against `dop/standard` may be calling a stale path.\n- The Seedance cross-reference for edit and extend means \"one Higgsfield video workflow\" can leave Cinema Studio without the docs shouting that the next call is a ByteDance model. Read the note before you attribute an edit to Cinema Studio.\n- Creative enums are long, and `auto` is an invalid value. Clients that send `auto` as a default will fail.\n- Outputs expire after the retention window. A pipeline that stores only the Higgsfield URL will break later.\n- Multi-shot prompting, in the Kling sense, is not a field. Pacing `single-shot` does not mean the other pacing values return an edit decision list.\n\n## Alternatives\n\n- [Seedance](/models/seedance) is the hosted model Cinema Studio's own note points at for edit and extend, and it has published per-second prices on fal and credit prices on Runway. Cinema Studio adds the camera and style enums and, on this schema, stops at 720p.\n- [Kling](/models/kling) has a public unit price, native-audio rows, and a multi-shot prompt field. It is also hosted inside Higgsfield as a different endpoint family.\n- [Runway Gen](/models/runway-gen) publishes 12 credits per second for Gen-4.5. It does not publish Cinema Studio's camera enum.\n- [Luma Ray](/models/luma-ray) is the closer cousin to Genjutsu when the job is modify or motion transfer, with a public dollar API and no native audio on fresh text generations.\n- [Veo](/models/veo) is the short Google clip with a GA card and a higher with-audio list price. It was not a model in the Higgsfield catalog index that was scanned.\n\n## Related tools\n\n- [/tools/higgsfield](/tools/higgsfield) is the product: plans, studio UI, and the MCP surface. Model rates on this page stay not observed even if the tool page discusses subscriptions.\n- [/tools/capcut](/tools/capcut) finishes captions, covers, and any upscale or reframing Cinema Studio's 720p cap does not cover.\n- [/tools/elevenlabs](/tools/elevenlabs) replaces generate-audio when the boolean is the wrong voice.\n- [/tools/seedance](/tools/seedance), [/tools/kling](/tools/kling), and [/tools/runway](/tools/runway) are the apps for the alternative models, including models Higgsfield also hosts.\n- [/tools/creatify](/tools/creatify) and [/tools/heygen](/tools/heygen) cover packaged ads and presenters.\n\n## FAQ\n\n### Is this page the Higgsfield app?\n\nNo. The app and its subscription live at [/tools/higgsfield](/tools/higgsfield). This page is Cinema Studio 4.0, Genjutsu, and the Soul still family as models, plus the explicit statement that hosted Kling or Seedance calls are not \"the Higgsfield model.\"\n\n### How much does one Cinema Studio clip cost?\n\nNot observed. The public schema and the billing chapter do not print a Cinema Studio credit rate. Call the account estimate endpoint. Ignore the Soul sample dollars on the billing page.\n\n### Does DoP still exist?\n\nNot in the `llms-full.txt` catalog saved on 2026-10-16. An indexed video guide still shows `POST /higgsfield-ai/dop/standard`. This environment could not save that guide as the guide URL. It redirected. Treat DoP as unverified until the console shows a current schema.\n\n### Can I generate 9:16 with audio?\n\nCinema Studio's schema allows aspect `9:16` and a generate-audio boolean. Resolution on that schema is `480p` or `720p`. Duration is 4 to 30 seconds. A price for that combination was not published in the files read here.\n\n### What is Genjutsu for?\n\nMotion transfer, object swap, and restyle of a source video. Motion transfer needs a source of at least 4 seconds, trims at 30, and wants 1 to 8 reference images. It is not the text-to-video path. Cinema Studio is the text-to-video path when you send no references.", "path": "/models/higgsfield" }, { "capabilities": [ "text-to-video", "image-to-video", "native-audio", "multi-shot" ], "verified_on": "2026-10-16", "native_audio": true, "multi_shot": true, "t2v": true, "i2v": true, "open_source": false, "in_launch_feed": true, "name": "Kling", "summary": "Kling 3.0 is a text-to-video and image-to-video family with native audio and multi-shot prompts on the cards opened on 2026-10-16. API units and fal seconds are the verified prices; membership dollars were not re-opened.", "example": false, "price_band": "Official API from $0.084 to $0.42 per second by tier on 2026-10-16; fal 3 Pro $0.112 to $0.196 per second", "where_to_run": [ "Kling developer API", "fal Kling 3 Pro text-to-video", "Runway API Kling credit rows" ], "slug": "kling", "markdown": "## Overview\n\nKling is Kuaishou's commercial video model family. This page selects among versions and hosts. It does not generate a clip, and it does not replace the product review at [/tools/kling](/tools/kling). The tool page is the app, the membership wallet, and the editor around the model. This page is the capability list, the places a job can be submitted, and the prices that were actually on a page on 2026-10-16.\n\nFilter tags: text-to-video, image-to-video, native audio, and multi-shot are all true. The developer homepage saved from `https://kling.ai/dev/pricing` says the video API supports text-to-video, image-to-video, multi-image-to-video, multi-element video editing, video extension, avatar, lip sync, video effects, text-to-audio, and video-to-audio. Separate links point at `/document-api/api/video/text-to-video` and `/document-api/api/video/image-to-video`. fal's Kling 3.0 Pro text-to-video card says the model has native audio and multi-shot support, and its schema includes `generate_audio` and `multi_prompt`. `open_source` is false. A weight file was not part of any page opened here.\n\nStatus of the 3.0 family: the public API price list is live, territory marked global on the saved developer HTML, with list prices in units. That is general availability for the priced 3.0, 3.0 Turbo, and 3.0 Omni rows. Kling 4.0 is a different question. The string \"4.0\" does not appear in the developer pricing HTML saved on 2026-10-16. A Kling blog URL returned a territory shell with no article body. Third-party posts disagree about whether 4.0 is generally available. This page does not adopt an early-access spec, a 30-second claim, or a credit rate for 4.0. Those items are not observed. Membership plan names and dollars on `kling.ai` were not re-opened in this pass, including any plan that the tool-page review distinguished from a third-party \"Ultra+\" label. Do not import those dollars from memory.\n\nRegion limits beyond the global territory flag on the developer page were not observed. A waitlist for 3.0 was not on the price JSON. A waitlist for 4.0 was not verified because the 4.0 article itself was not saved.\n\n## Specs\n\nVersion line. Price the 3.0 generation that is on the API card. Treat 2.5 and O3 rows on Runway's credit guide as host SKUs, not as a reason to describe 4.0. Omni is on the official unit table as its own model block.\n\nOfficial unit table, video generation API, subtitle \"Video API Pricing 1 Unit = $0.14 (list price)\":\n\n| Model | Function | 720p | 1080p | 4K |\n| --- | --- | --- | --- | --- |\n| Kling 3.0 Turbo | With native audio | 0.8 units ($0.112)/s | 1.0 unit ($0.14)/s | not offered (`-`) |\n| Kling 3.0 | No native audio | 0.6 units ($0.084)/s | 0.8 units ($0.112)/s | 3.0 units ($0.42)/s |\n| Kling 3.0 | With native audio, no voice control | 0.9 units ($0.126)/s | 1.2 units ($0.168)/s | 3.0 units ($0.42)/s |\n| Kling 3.0 Omni | No video input, no native audio | 0.6 units ($0.084)/s | 0.8 units ($0.112)/s | 3.0 units ($0.42)/s |\n| Kling 3.0 Omni | No video input, with native audio | 0.8 units ($0.112)/s | 1.0 unit ($0.14)/s | 3.0 units ($0.42)/s |\n| Kling 3.0 Omni | With video input, no native audio | 0.9 units ($0.126)/s | 1.2 units ($0.168)/s | 3.0 units ($0.42)/s |\n\nOther model rows on the same card: Motion Control 720p 0.9 units ($0.126)/s and 1080p 1.2 units ($0.168)/s. Avatar 720p 0.4 units ($0.056)/s and 1080p 0.8 units ($0.112)/s. Effects pricing is a link, not a single rate. The dollar figures in parentheses are printed by Kling next to the unit counts. They are not a conversion this directory invented. Resource-pack prices, including any pack that excludes 4.0, were not in this HTML. They are not observed for this pass.\n\nfal Kling Video v3 Pro text-to-video, model id `fal-ai/kling-video/v3/pro/text-to-video`, endpoint `https://fal.run/fal-ai/kling-video/v3/pro/text-to-video`, read in full on 2026-10-16:\n\n| Field | Observed value |\n| --- | --- |\n| Duration | `\"3\"` through `\"15\"`. Default `\"5\"`. |\n| Aspect | `16:9`, `9:16`, `1:1`. Default `16:9`. |\n| Audio | `generate_audio` default true. Chinese and English voice output. Other languages are translated to English. English speech wants lowercase letters. Acronyms and proper nouns want uppercase. |\n| Multi-shot | `multi_prompt` overrides the single prompt and divides the video into shots. Provide prompt or `multi_prompt`, not both. `shot_type` is `customize` (default) or `intelligent`. Intelligent lets the model determine shot structure. |\n| Guidance | `cfg_scale` default 0.5, range 0 to 1. |\n| Negative prompt | Default \"blur, distort, and low quality\". |\n| Resolution | Not a field on this Pro text-to-video schema. |\n\nOutput is a video file. The example container is mp4. A 4K switch, a frame-rate field, and a seed field were not on this schema. Resolution dollars come from Kling's unit table, not from fal's Pro card. An image-to-video fal schema was not the file saved. Image-to-video is still a documented official API route, so the filter stays true, and the fal field list above is text-to-video only.\n\nMaximum reference images, the Omni element cap, and a 4.0 keyframe count were not on the saved price JSON. A marketing sentence on the developer page uses the words \"coherent multi-shot narrative\" inside a sample prompt. That sentence is not a numeric spec. The numeric multi-shot control that was actually parsed is fal's `multi_prompt`.\n\n### Prompt notes\n\nfal's published example, not a clip this directory rendered: \"Close-up of glowing fireflies dancing in a dark forest at twilight. Soft bioluminescent particles float through the air. Shallow depth of field, bokeh lights in background. Magical atmosphere, gentle movement.\" It is one shot, one mood, one camera distance. If you need several shots, the schema says to switch to `multi_prompt` and to stop sending the single `prompt` at the same time. The voice rule is operational: lowercase for ordinary English speech, uppercase for names and acronyms. Whether that rule survives a host that rewrites the prompt was not tested.\n\nArena scores were not opened. They are not observed.\n\n## Where to run\n\n1. Kling's own developer price page: [kling.ai/dev/pricing](https://kling.ai/dev/pricing). This is the unit table. The API index links text-to-video and image-to-video documents under `/document-api/api/video/`.\n2. fal's Pro text-to-video card: [fal-ai/kling-video/v3/pro/text-to-video](https://fal.ai/models/fal-ai/kling-video/v3/pro/text-to-video). This is the duration enum, the aspect enum, and the audio and voice-control prices.\n3. Runway's developer guide: [Runway API pricing](https://docs.dev.runwayml.com/guides/pricing). Kling model ids there are host SKUs billed in Runway credits. The Runway product is [/tools/runway](/tools/runway).\n4. The Kling app itself is [/tools/kling](/tools/kling). Membership credit math was not re-fetched, so this model page does not pretend the app price and the API unit price are one number.\n\nHiggsfield's saved API catalog includes many Kling 3 endpoints (standard, pro, turbo, 4K, text and image, plus motion control and Omni reference routes). That catalog is a fourth door. Its per-call dollar price was not in the schema extract. See [/tools/higgsfield](/tools/higgsfield) for the studio, and do not confuse those hosted Kling routes with Higgsfield's own Cinema Studio model at [/models/higgsfield](/models/higgsfield).\n\n## Multi-channel pricing\n\nVerified 2026-10-16.\n\n**Official API, unit list price.** 1 unit = $0.14. The table in Specs is the channel. A 5-second 1080p Turbo clip with native audio is 5 times $0.14 = $0.70 if the printed per-second figure is applied with no other adder. That product is derived. Failed generations and concurrency were not given a dollar rule in the extracted video table. Image rows on the same page are a different meter: Kling Image 3.0 is 8 units ($0.028) per image at 1K or 2K. Do not mix an image unit with a video second.\n\n**fal, Kling 3 Pro text-to-video.** $0.112 per second with audio off, $0.168 per second with audio on, $0.196 per second with voice control. fal's own example: a 5-second video with audio on and voice control costs $0.98. Resolution is not part of this price. You cannot put fal's $0.168 next to official 4K $0.42 and call them the same SKU. The official 1080p \"with native audio, no voice control\" cell is also $0.168, and the 720p \"no native audio\" cell is $0.084. The match at $0.168 is a coincidence to notice, not a contract that fal Pro equals official 1080p.\n\n**Runway credits, $0.01 per credit on the developer guide.** Printed video rows include `kling3.0_standard` 9 credits per second without audio and 13 with audio; `kling3.0_pro` 12 without audio and 17 with audio; `kling3.0_4k` 43 credits per second; `klingO3_standard` 13; `klingO3_pro` 17; `klingO3_4k` 43; `kling2.5_turbo_pro` 12. Derived dollars at $0.01: standard silent $0.09/s, standard audio $0.13/s, pro silent $0.12/s, pro audio $0.17/s, 4K $0.43/s. Label them derived. They are close to Kling's own unit dollars and they are not the same invoice. Luma's pricing page also lists Kling 3.0 and Kling Omni in Luma credits (for example 30 credits per second at 720p without the audio Yes, 40 with audio Yes). That third-party credit column was on [lumalabs.ai/pricing](https://lumalabs.ai/pricing) on 2026-10-16. A Luma credit's dollar value was not printed on the Kling rows.\n\nMembership packs and the in-app credit burn for 3.0 versus 4.0 are not observed in this pass. If the tool page states plan prices from an earlier same-day read, treat them as tool-page notes until the membership URL is opened again.\n\n## Workflow and platform fit\n\nWorkflow names are site labels. They are not live guide URLs.\n\n**Faceless.** Kling 3 Pro on fal can be asked for 9:16 and for 3 to 15 seconds with native audio. That covers a faceless hook and a short scene. A 60-second narration still needs an editor. Captions are not a Kling output field. Finish them in [/tools/capcut](/tools/capcut). If the voice must be a specific cloned speaker, the fal note says non-English speech is translated to English. That is a poor fit for a non-English faceless channel unless you cover the clip with [/tools/elevenlabs](/tools/elevenlabs) and turn generation audio off. The official price table has a cheaper no-audio row for that path.\n\n**UGC ads.** Image-to-video and motion control are the productized ad routes. Motion control is priced on the official card. A full UGC template system is closer to [/tools/creatify](/tools/creatify) than to a raw Kling prompt. Use Kling when you already have the first frame and the motion reference.\n\n**AI avatar.** Kling's price card has an Avatar row, separate from the general 3.0 video rows, at $0.056/s in 720p and $0.112/s in 1080p. That is a Kling avatar meter, not HeyGen's credit meter. For a photo presenter with a documented per-second enterprise rate, compare [/models/heygen-video](/models/heygen-video). For a cinematic human who is not locked to a avatar rig, Kling 3 or Seedance is the usual short list.\n\n**Long-to-short.** Kling does not take a one-hour source and return chapters. Video extension is named as a supported API feature. A maximum extend length was not in the price JSON. Cut the long source in CapCut, then extend or restyle a short beat if the API you call documents extension.\n\n**Captions and cover.** Not part of the generation schema that was read. Keep them in the editor. A cover image can be a Kling image generation, which is the per-image unit price above, then typeset in CapCut.\n\n**Voice and publish.** Native audio is a priced mode, not a free toggle, on both the official 3.0 rows and the fal Pro card. Turbo's only audio row is \"with native audio.\" A silent Turbo price was a dash at 4K and was not given as a separate 720p silent cell. Publish 9:16 when the fal aspect is set to `9:16`. Official unit prices do not split by aspect. Whether 9:16 changes the unit rate was not stated. Assume the second price stands until a document says otherwise.\n\n**Shorts vertical fit.** The fal card is explicit: `9:16` is an option, and 15 seconds is inside the duration enum, which matches a short vertical ad. 4K is an official 3.0 tier at $0.42/s and is usually wasted on a phone Short unless a brand master is the deliverable. 720p or 1080p is the practical Shorts band. Audio translation to English is the main vertical-channel risk for non-English speech.\n\n## Pros\n\n- The official card prints both units and dollars, so a reader can audit 0.8 units times $0.14 without trusting a blog screenshot.\n- Multi-shot is a schema, not a slogan, on fal 3 Pro: `multi_prompt` plus `shot_type` of customize or intelligent, with an exclusive-or against the single prompt.\n- 9:16, 16:9, and 1:1 are on the fal Pro card, which is enough for Shorts, a landscape master, and a square cut from the same request shape.\n- Silent and with-audio prices are separate on official 3.0 and on fal, so a B-roll job does not have to pay the speech rate.\n- Image-to-video, motion control, and avatar are separately named and, for motion control and avatar, separately priced. You can refuse the features you will not call.\n\nThese are readings of opened pages, not a lab score.\n\n## Cons\n\n- Resolution is missing from the fal Pro text-to-video schema, so a fal invoice and a Kling 720p versus 1080p choice are not the same decision. Quote both channels or you will understate 4K.\n- Kling 4.0's API price, duration, and reference caps were not on the saved developer HTML. Repeating a search snippet as if it were the price page would be a false spec.\n- Membership credit burn was not re-opened. An app \"credit\" and an API \"unit\" are different words. The tool page is the place that discussed plans. This page refuses to restate plan dollars it did not fetch.\n- fal's language note translates other languages to English for voice output. A multilingual Shorts system can fail quietly if someone assumes the prompt language is the speech language.\n- The developer page's sample \"multi-shot\" prompt is marketing copy. Only `multi_prompt` should be cited as the control.\n- Resource packs and volume discounts were not in the parsed price JSON. A custom pack is \"contact us\" on the page, with no public rate.\n\n## Alternatives\n\n- [Seedance](/models/seedance). Seedance 2.5 on fal runs 4 to 30 seconds and does not charge extra for audio on the US text card. Kling's advantage on the opened schemas is the explicit multi-shot list and the official 4K unit price.\n- [Veo](/models/veo). Veo 3.1's GA length enum is 4, 6, or 8 seconds. Kling's fal enum is every integer from 3 to 15. Veo's Gemini list price at 720p/1080p with audio is $0.40 per second, above Kling's printed 1080p audio rates.\n- [Runway Gen](/models/runway-gen). Gen-4.5 is 12 Runway credits per second with no audio split on that row. Kling on Runway is the `kling3.0_*` rows. Pick Gen when you want Runway's own model and the professional-output surcharge. Pick Kling when you want the unit table above.\n- [MiniMax Hailuo](/models/minimax-hailuo). Official Hailuo 3 is $0.08 per second at 768P and $0.13 at 2K, which is a lower list price than Kling 1080p with audio. Hailuo's official page did not describe generated native audio. Kling did.\n- [Luma Ray](/models/luma-ray). Ray3.2 has no native audio on text-to-video or image-to-video, and it does have 16 keyframes, HDR, and EXR. Choose Ray for a grade. Choose Kling for speech in the same request.\n\nNo `/models/kling-vs-veo` page exists yet. The sentences above are the comparison.\n\n## Related tools\n\n- [/tools/kling](/tools/kling) is the product. Use it for the app path and for membership notes that were checked on the tool pass.\n- [/tools/runway](/tools/runway) resells Kling SKUs in credits. Match the model id before you compare a credit to a Kling unit.\n- [/tools/higgsfield](/tools/higgsfield) lists Kling 3 routes in its API catalog. Those routes are hosted Kling, not the DoP or Cinema Studio models.\n- [/tools/capcut](/tools/capcut) is the caption, cover, and vertical-finish step after the clip returns.\n- [/tools/elevenlabs](/tools/elevenlabs) replaces or covers Kling voice when you cannot accept the English-translation rule.\n- [/tools/creatify](/tools/creatify) is an ad-template product when the job is a UGC structure rather than a single Kling shot.\n- [/tools/heygen](/tools/heygen) is the presenter product when the Kling Avatar row is the wrong rig.\n\n## FAQ\n\n### Is Kling 4.0 generally available on the API?\n\nNot observed. The developer pricing HTML saved on 2026-10-16 prices 3.0, 3.0 Turbo, and 3.0 Omni. It does not contain a 4.0 row. A blog fetch did not return an article body. Do not quote unpublished 4.0 credit rates.\n\n### Why do fal and the official site both show $0.168?\n\nfal's Pro text-to-video card charges $0.168 per second with audio on, and the official 3.0 1080p \"with native audio, no voice control\" cell also prints $0.168. fal's card has no resolution field. Treat them as two published numbers that match, not as proof of one SKU.\n\n### Can one request contain several shots?\n\nOn fal's 3 Pro text-to-video schema, yes, by sending `multi_prompt` and not also sending `prompt`. `shot_type` can be `customize` or `intelligent`. A maximum shot count was not stated on that schema's field text. Duration of the whole video is still one of 3 through 15 seconds.\n\n### Does 9:16 cost more?\n\nThe official unit table is headed by resolution, not by aspect. The fal Pro price is per second and does not mention aspect. A vertical upcharge was not observed.\n\n### Were membership prices checked today?\n\nThe API unit page and the fal card were checked on 2026-10-16. The membership plan page was not re-opened for this model pass. Plan dollars are not observed here. Use the tool page only as a separate note, and re-check the membership URL before you publish a plan figure.", "path": "/models/kling" }, { "capabilities": [ "text-to-video", "image-to-video", "video-to-video", "multi-keyframe" ], "verified_on": "2026-10-16", "native_audio": false, "multi_shot": true, "t2v": true, "i2v": true, "open_source": false, "in_launch_feed": true, "name": "Luma Ray", "summary": "Ray3.2 is Luma's current video model for text-to-video, image-to-video, modify, and reframe, with up to 16 keyframes and no native audio on new generations. App credits and API dollars were both on Luma's pages on 2026-10-16.", "example": false, "price_band": "API SDR from $0.15 to $1.20 per 5-second T2V or I2V clip by resolution; Plus plan $30 per month on 2026-10-16", "where_to_run": [ "Luma Ray app", "Luma Ray3.2 API", "Luma plans that include third-party models" ], "slug": "luma-ray", "markdown": "## Overview\n\nRay3.2 is the video model Luma tells developers to integrate, on both the public Ray page and the API page fetched on 2026-10-16. Luma's own `llm-info` material in the same research pass described Ray3.2 as the current model and treated older Dream Machine and Ray2 names as deprecated. This page uses Ray3.2 as the current line and mentions Ray3.14 only where the pricing page still prints a credit table for it. There is no separate Luma tool page in the star set. The doors are Luma's site and API. This directory does not generate Ray clips.\n\nFilter tags: text-to-video and image-to-video are true. The Ray page prices both, and the FAQ calls them text2video and image2video. Native audio is false. The FAQ answer \"Do generations have native audio?\" says text2video and image2video have no natively generated audio, so voiceover, music, and sound effects are a separate step. Modify Video and Reframe preserve original audio. That preservation is not generated speech. `multi_shot` is true because Multi-Keyframe is a documented control, up to 16 keyframes inside a single clip, including up to 16 keyframes for image-to-video. That is keyframe direction, not a Kling-style list of shot prompts. The filter uses the multi-shot bit because the information architecture treats multi-shot and multi-keyframe control as the same family of \"more than one beat in a request.\" The body is the precise meaning. `open_source` is false.\n\nStatus: Ray3.2 is on public pricing, on the Ray marketing page, and on the API cost table. That is general availability for buyers of those plans or of pay-per-video API calls. A waitlist was not on those pages. A region lock was not on those pages. Team and Enterprise are \"contact us,\" which is a sales motion, not a closed beta for the model. The API scale tier mentions dedicated capacity, a minimum of 4 units, an SLA, and a no-train guarantee. One unit there is defined as 1 request per minute on the base tier or 0.4 RPM on the max tier. That sentence is capacity packaging, not the per-video price.\n\n## Specs\n\nVersion line. Ray3.2 is current. Ray3.14 still has rows on the pricing page: text-to-video and image-to-video at 4 credits per second draft, 10 at 540p, 20 at 720p, and 80 at 1080p, with HDR at 4× on 3.14 versus 2× on 3.2. Use 3.14 only if you are maintaining an older integration. New selection should start at 3.2. The FAQ answer \"What Is Ray3.2?\" says: \"Ray3.2 is our newest and most versatile video model. It powers four workflows in one: Modify Video, image2video, text2video, and Reframe.\" Dream Machine is not the current name to brief.\n\nRay page and FAQ, fetched 2026-10-16:\n\n| Field | Observed value |\n| --- | --- |\n| Resolutions | 360p draft, 540p, 720p (default), 1080p. |\n| Text and image length | 5 or 10 second native clips. |\n| Modify Video | Up to 20 seconds, depending on frame rate. 24 fps up to 20 seconds, 30 fps up to 15 seconds, 60 fps up to 7 seconds. |\n| Reframe length | One FAQ answer says source videos up to 16 seconds. Another answer on the same page says up to 12 seconds. Both are printed. This page does not pick a winner. |\n| Keyframes | Up to 16 in a clip, including image-to-video. |\n| Aspects | 9:16, 3:4, 1:1, 4:3, 16:9, 21:9. |\n| Audio | No native audio on new text or image generations. Modify and Reframe keep source audio. |\n| Container | Standard output mp4. Also native 16-bit HDR and 16-bit EXR (ACES2065-1 / AP0) for grading. |\n| Modify inputs | mp4, mov, wmv. |\n| Faces | Expressive facial performance for up to 8 faces, when that Modify feature is on. Lip sync preserved when face tracking is on, per the FAQ. |\n\nAPI page: \"1080p output across the model. V2V at up to 20 seconds.\" HDR generation and 16-bit EXR export are repeated there. Draft 360p is on the credit table and is missing from the API dollar grid, which starts at 540p.\n\nCredit table on the Ray page for Ray3.2, SDR base:\n\n| Task | 360p draft | 540p | 720p | 1080p |\n| --- | --- | --- | --- | --- |\n| T2V/I2V 5s | 20 | 50 | 100 | 400 |\n| T2V/I2V 10s | 60 | 150 | 300 | 1200 |\n| V2V 5s | 180 | 240 | 360 | 720 |\n| V2V 10s | 360 | 480 | 720 | 1440 |\n| Reframe per second | 10 | 20 | 40 | 120 |\n\nHDR is 2× SDR credits. HDR plus EXR is 3×. The audio column on the pricing-page Ray3.2 block prints a dash, which matches the FAQ.\n\nElevenLabs is not a Ray audio feature. If a Luma screen offers an ElevenLabs connection, that is a separate product: [/tools/elevenlabs](/tools/elevenlabs). It was not a line on the Ray price table read here.\n\n### Prompt notes\n\nThe Ray page describes controls more than it gives a single canonical prompt. For Modify, the FAQ separates Motion (body, choreography, camera dynamics, which the tool tries to preserve) from Structure (shapes, edges, composition). A prompt that fights the source motion will not behave like a fresh text-to-video prompt. The page says Modify struggles with heavy camera shake, whip-fast action, blur, darkness, low resolution, crowded overlapping subjects, and occluded faces. That limitation is the vendor's own FAQ, not a lab note from this directory. No clip was generated here.\n\nArena scores were not opened. Not observed.\n\n## Where to run\n\n1. The Ray product page: [lumalabs.ai/ray](https://lumalabs.ai/ray). Specs, keyframes, credit table, and the FAQ JSON that answers the audio question.\n2. The API cost page: [lumalabs.ai/api](https://lumalabs.ai/api). Pay-per-video SDR dollars, HDR multipliers, and the statement that there is no minimum commitment on that pay-per-video tier.\n3. Plan cards: [lumalabs.ai/pricing](https://lumalabs.ai/pricing). Plus, Pro, and Ultra credit pools, plus the same Ray3.2 credit math and rows for other companies' models.\n\nThe API page says partners include a path such as Fal. A fal Ray3.2 `llms.txt` was not saved in this pass, so a fal dollar card for Ray is not observed. Do not invent one. Luma's own API table is the dollar channel.\n\n## Multi-channel pricing\n\nTwo Luma channels were opened on 2026-10-16: subscription credits and API dollars. They are both official. A third-party host price was not observed.\n\n**API, SDR, pay per video, no minimum commitment.** Text-to-video and image-to-video, 5 seconds: $0.15 at 540p, $0.30 at 720p, $1.20 at 1080p. Ten seconds: $0.45, $0.90, and $3.60. Video-to-video, 5 seconds: $0.72, $1.44, and $2.16. Ten seconds: $1.08, $2.16, and $4.32. Reframe per second: $0.06, $0.12, and $0.36 at 540p, 720p, and 1080p. The page says prices shown are SDR. HDR is 2× those figures. HDR plus EXR is 3×. A 1080p 5-second HDR text clip is $1.20 times 2 = $2.40 derived. Draft 360p has credits and does not have a row in this dollar grid. Rate limits apply and that pay-per-video tier has no latency SLA, per the page. The scale tier is contact sales.\n\n**App credits, Ray page and pricing page.** The table in Specs. Plans that buy the credits: Plus is $30 per month or $300 billed yearly, 10,000 credits, commercial use, Luma and third-party models. Pro is $90 per month or $900 billed yearly, 40,000 credits, and 4× Luma Agents usage. Ultra is $300 per month or $3,000 billed yearly, 150,000 credits, and 15× agents. The pricing metadata says save up to 20% with yearly. The billed-yearly totals are not twelve times the monthly price: Plus $300 versus $30 × 12 = $360, Pro $900 versus $90 × 12 = $1,080, and Ultra $3,000 versus $300 × 12 = $3,600. Each is about 17% less. The plan data on that page also includes `yearlyPrice` values of $25, $75, and $250, which are those yearly totals divided by 12. None of the three is a full 20% off. The \"up to 20%\" line is the vendor headline. No free plan was on the cards that were extracted. Team and Enterprise are contact.\n\nA worked credit check, derived: one 720p 5-second text clip is 100 credits. Plus's 10,000 credits cover 100 such clips if nothing else is spent and HDR is off. One 1080p 5-second text clip is 400 credits, so the same pool covers 25. A 1080p HDR clip is 800 credits. These products assume the table and no failed-generation rebate. A rebate rule was not in the extracted FAQ answers.\n\nThird-party models on the same pricing page, including Seedance, Kling, Veo, and MiniMax, spend the same credit pool and are not Ray. Their rows belong on their own model pages. Mixing a Ray 100-credit clip with a Seedance per-second credit row is how teams blow the pool.\n\n## Workflow and platform fit\n\nLabels only. Real workflow guides for these names are not linked.\n\n**Faceless.** Ray3.2 can do 9:16 and a 5 or 10 second native clip, which is a B-roll beat. It will not speak. Every faceless cut needs an external voice, then captions. Use [/tools/elevenlabs](/tools/elevenlabs) and [/tools/capcut](/tools/capcut). Modify can restyle footage you already shot and keep its audio, which is a different faceless pattern: film once, restyle many times, keep the take's sound.\n\n**UGC ads.** Product swap and environment change are named on the Ray page as workflows. They are marketing names for Modify-style jobs, not a separate price SKU in the table. A testimonial face that must stay locked is a weaker pitch than HeyGen unless your test shows Modify's face tracking holds. The vendor says face tracking preserves lip sync. It also lists occluded faces as a failure mode. [/tools/creatify](/tools/creatify) remains the template tool for a structured UGC ad.\n\n**AI avatar.** Ray is not an avatar engine. Up to 8 faces in Modify is performance transfer, not a photo-to-spokesperson API. See [/models/heygen-video](/models/heygen-video).\n\n**Long-to-short.** Modify's maximum is on the order of 7 to 20 seconds depending on frame rate, and reframe is 12 or 16 seconds depending on which FAQ line you cite. A long episode must be cut first. Ray does not replace a long-to-short editor.\n\n**Captions and cover.** No caption model is in the Ray spec. A cover can be a frame from the mp4 or a still from Uni-1, which the API page sells as a separate image model. Uni-1 prices were on the API page and are out of scope for this video sheet beyond that mention.\n\n**Voice and publish.** Publish path is silent picture plus a voice track, or Modify of a clip that already has a voice. 9:16 is in the aspect list, so Shorts, Reels, and TikTok frames are specified. HDR and EXR are for a grade in Resolve or Nuke, which the API page names. They are the wrong master to upload untouched to a phone feed. Export or transcode to a normal mp4 before you schedule the Short.\n\n**Shorts vertical fit.** Strong on aspect and on the existence of a draft 360p credit tier for cheap previz. Weak on audio and on length if the Short needs a continuous 20-second generated scene from text. Text and image generations stop at 10 seconds. A vertical ad cut from a modified source can run longer, up to the frame-rate cap, and can keep the performer's audio. That is the Shorts pattern Ray actually documents.\n\n## Pros\n\n- The audio question has a direct FAQ answer, so you do not have to infer silence from a dash in a table. The dash and the answer agree.\n- Sixteen keyframes is a concrete control, repeated for image-to-video and for Modify.\n- 9:16 is in the aspect list, alongside a cinematic 21:9, so one model card covers a Short and a wide master.\n- HDR 2× and HDR+EXR 3× are written next to the credit table, and the API repeats the same multipliers in dollars.\n- App credits and API dollars were both public on the same day, so a studio can compare a subscription pool with pay-per-video without a sales call.\n- Modify's failure modes are published by the vendor, which is more useful than a feature grid that only lists wins.\n\n## Cons\n\n- New text and image clips have no native audio. Any workflow that wanted one-request speech should look at Kling, Seedance, or Veo instead.\n- Text and image generations are only 5 or 10 seconds. The longer numbers belong to Modify and depend on source frame rate.\n- Reframe's maximum length is stated as both 16 seconds and 12 seconds on the same page. Quote both until Luma edits one of them.\n- Draft 360p is missing from the API dollar grid. A developer cannot assume the 20-credit draft maps to a published dollar.\n- 1080p text-to-video is $1.20 for 5 seconds SDR, before HDR. That is a different budget from 540p at $0.15. The cheap headline and the master price should not be swapped.\n- Ray3.14 rows are still on the pricing page. An old integration can bill the 4× HDR multiplier by mistake.\n- No non-Luma price card was saved. Host arbitrage for Ray is not observed.\n\n## Alternatives\n\n- [Seedance](/models/seedance) generates audio inside the fal price and runs up to 30 seconds on the 2.5 cards. Ray is the one with keyframes, modify, and EXR.\n- [Kling](/models/kling) has native audio, a multi-shot prompt list, and an official unit table. It does not publish a 16-keyframe Modify FAQ of the kind Ray does.\n- [Veo](/models/veo) has sound generation, an 8-second cap, and a GA card in `us-central1`. Ray's pay-per-video API is the easier public dollar meter. Veo's is split between preview Gemini prices and a Vertex card without pay-as-you-go.\n- [Runway Gen](/models/runway-gen) is 12 Runway credits per second with an HDR surcharge and no confirmed 9:16 enum on its price guide. Ray's 9:16 is explicit.\n- [MiniMax Hailuo](/models/minimax-hailuo) is priced per second on MiniMax's own pay-as-you-go page, without Ray's keyframe and EXR story.\n\n## Related tools\n\n- [/tools/elevenlabs](/tools/elevenlabs) is the voice step the Ray FAQ tells you to add.\n- [/tools/capcut](/tools/capcut) holds captions, covers, and the cut that brings Ray beats under one timeline.\n- [/tools/runway](/tools/runway), [/tools/kling](/tools/kling), and [/tools/seedance](/tools/seedance) are the other studios when you leave Luma's credit pool.\n- [/tools/higgsfield](/tools/higgsfield) hosts several of those alternative models. A Ray endpoint was not in the Higgsfield catalog snapshot read for this pass.\n- [/tools/heygen](/tools/heygen) and [/tools/creatify](/tools/creatify) cover presenters and ad templates that Ray's Modify path does not productize as an avatar.\n\nLuma's own app is the primary tool and it does not have a `/tools` slug in this directory yet. The three links above satisfy the related-tool rule. The missing Luma tool page is a directory gap, not a claim that Ray has no app.\n\n## FAQ\n\n### Does Ray3.2 generate audio?\n\nNot for text-to-video or image-to-video. The official FAQ says there is no natively generated audio on those modes. Modify and Reframe keep the audio that was already in the source file.\n\n### How long is a clip?\n\nText-to-video and image-to-video: 5 or 10 seconds. Modify: up to 20 seconds at 24 fps, 15 at 30 fps, and 7 at 60 fps. Reframe: the page says both 12 and 16 seconds in different answers.\n\n### How many keyframes?\n\nUp to 16, including image-to-video, according to the FAQ.\n\n### What does a 1080p clip cost on the API?\n\nA 5-second SDR text or image clip at 1080p is $1.20 on the API table fetched 2026-10-16. HDR doubles that. HDR plus EXR triples it. The same clip in the app is 400 credits before multipliers.\n\n### Is Ray open source?\n\nNo weight release was on the pages opened. `open_source` is false. The API is a paid service with a separate scale tier for an SLA.", "path": "/models/luma-ray" }, { "capabilities": [ "text-to-video", "image-to-video" ], "verified_on": "2026-10-16", "native_audio": false, "multi_shot": false, "t2v": true, "i2v": true, "open_source": false, "in_launch_feed": true, "name": "MiniMax Hailuo", "summary": "MiniMax-H3, also called Hailuo 3, is the current pay-as-you-go video model on MiniMax's official price page, with legacy Hailuo 2.3 and 02 clip prices on the same page. Generated native audio was not described there.", "example": false, "price_band": "Official MiniMax-H3 $0.05 to $0.13 per second by variant and resolution on 2026-10-16", "where_to_run": [ "MiniMax official pay-as-you-go API", "fal Hailuo 2.3 and Hailuo-02 text-to-video", "Higgsfield MiniMax H3 endpoints", "Hailuo app" ], "slug": "minimax-hailuo", "markdown": "## Overview\n\nHailuo is MiniMax's video model line. On the official pricing overview fetched on 2026-10-16, the current per-second video models are MiniMax-H3 and MiniMax-H3-Max. The same page still lists clip prices for MiniMax-Hailuo-2.3, Hailuo-2.3-Fast, and Hailuo-02. This selection page uses H3 as the current meter and keeps the legacy clip prices because they are on the official page, not because they are the model you should default to. The older name Hailuo and the newer name H3 refer to one family. Say which id you are billing.\n\nThis is not a MiniMax app review. The 2026-10-16 pass for this page did not open the consumer subscribe page. The Hailuo app is [/tools/hailuo](/tools/hailuo), verified 2026-10-17. Yearly cards and the MiniMax API rates from that later check live on the tool page. Hosts and the official API remain the doors for the rates in the tables below.\n\nFilter tags: text-to-video and image-to-video are true for H3 because Higgsfield's saved API catalog documents `POST /minimax/h3/text-to-video` and `POST /minimax/h3/image-to-video`, and fal documents text-to-video for Hailuo 2.3 standard and Hailuo-02 standard. The official price table itself does not label H3 rows as text versus image. The host schemas do. Native audio is false. The official page lists Audio as an input material, and that input is Free. Free input audio is not a description of generated speech or generated sound effects. Luma's pricing table prints Yes in an Audio column for Minimax H3. That column is not defined on the extract as \"the model synthesizes a soundtrack.\" This page does not flip `native_audio` to true on the strength of that Yes. `multi_shot` is false. No multi-shot field was in the H3 schema extract. `open_source` is false. A weight release was not on the official price page. Third-party claims about open weights were not opened and are not adopted.\n\nStatus: the pay-as-you-go prices are public on `https://platform.minimax.io/docs/pricing/overview`, which is where `https://platform.minimax.io/docs/guides/pricing-paygo` redirected, HTTP 200, on 2026-10-16. That is general availability of the priced API, subject to whatever account MiniMax requires. A waitlist and a region block were not on the extracted price sections. A page last-updated line was not confirmed in the saved HTML. The verification date is the fetch date, 2026-10-16. Do not invent a September publish date from an earlier search snippet.\n\n## Specs\n\nVersion line. Prefer MiniMax-H3 or H3-Max for new spend. Keep Hailuo 2.3 and 02 in mind when a host still routes those ids. fal's `fal-ai/minimax/video-01-live` card is an older model at $0.50 per video. It was downloaded by mistake while looking for an H3 path. It is not H3. `fal-ai/minimax-h3/text-to-video/llms.txt` returned HTTP 404. The live fal cards used here are Hailuo 2.3 and Hailuo-02.\n\nOfficial output list prices:\n\n| Model | Resolution | List price |\n| --- | --- | --- |\n| MiniMax-H3 | 768P | $0.08 per second |\n| MiniMax-H3 | 2K | $0.13 per second |\n| MiniMax-H3-Max | 480P | $0.05 per second |\n| MiniMax-H3-Max | 768P | $0.08 per second |\n\nOfficial input materials:\n\n| Model | Material | Rule |\n| --- | --- | --- |\n| MiniMax-H3 | Audio | Free |\n| MiniMax-H3 | Image | First 5 images free, then $0.04 each |\n| MiniMax-H3 | Video | Billed by input duration and output resolution: 2K $0.13/s, 768P $0.08/s |\n| MiniMax-H3-Max | Audio | Free |\n| MiniMax-H3-Max | Image | First 2 images free, then $0.074 each |\n| MiniMax-H3-Max | Video | 480P $0.0553/s, 768P $0.143/s |\n\nRegeneration: MiniMax-H3-Regeneration from 768P to 2K is $0.05 per second of regenerated output. Regeneration inputs are billed again. Audio is free. The first 5 images are free, then $0.025 each. Input video from the original 768P task is $0.05 per second. H3-Context-IR is $0.90 per million input tokens and $3.60 per million output tokens. That row is tokens, not video seconds. Do not add it to a per-second quote.\n\nDuration, frame rate, aspect ratio, and a native output-audio switch were not on the official price table. Not observed there.\n\nHiggsfield's H3 text-to-video schema, from the `llms-full.txt` snapshot saved 2026-10-16, fills some of those holes for that host only. Duration is an integer from 5 to 15, default 5. Resolution is fixed to the case-sensitive value `2K`. Aspect values are `auto`, `adaptive`, `21:9`, `16:9`, `4:3`, `1:1`, `3:4`, and `9:16`. For text-only requests, `auto` or `adaptive` resolves to 16:9. The prompt must contain non-whitespace text, and the note says the provider serializer uses the first 7,000 characters. A boolean on the schema is the AIGC watermark, default false, not a generate-audio flag. Image-to-video uses the same 5 to 15 second range and the same fixed `2K` resolution. The prompt is required. `end_image_url` is an optional last frame. The output aspect follows the input keyframe even if `aspect_ratio` is sent. Reference-to-video is a third endpoint in the catalog. Its full field list was not copied here.\n\nfal Hailuo 2.3 standard text-to-video, `https://fal.run/fal-ai/minimax/hailuo-2.3/standard/text-to-video`: \"Standard, 768p.\" Duration options `\"6\"` and `\"10\"`, default `\"6\"`. Price $0.28 per 6 second video and $0.56 per 10 second video, which matches the official legacy 768P rows. `prompt_optimizer` default true. No resolution enum and no aspect enum were in the portion that defines those durations. The vendor example prompt is a storm over a city. It was not run here.\n\nfal Hailuo-02 standard text-to-video: \"$0.045/sec. A single 6 second video will cost $0.27/video.\" Duration options include 6 and 10, default 6, and \"10 seconds videos are not supported for 1080p resolution.\" The official Hailuo-02 768P 6-second price is $0.28, not $0.27. Both numbers are published. They are not averaged.\n\nLegacy official clip prices, same page: Hailuo-2.3-Fast is $0.19 per 768P 6s, $0.32 per 768P 10s, and $0.33 per 1080P 6s. Hailuo-2.3 and Hailuo-02 share $0.28 per 768P 6s, $0.56 per 768P 10s, and $0.49 per 1080P 6s. Hailuo-02 also lists $0.10 per 512P 6s and $0.15 per 512P 10s. A 1080P 10-second official row was not in that list. The fal note that 10 seconds is unavailable at 1080p on Hailuo-02 is consistent with the missing official 1080P 10s price, and it is still a fal sentence rather than an official duration policy for H3.\n\nSpeech prices on the same official page, such as speech-2.6-turbo, are not video rates. Image-01 at $0.0035 per image is not a video frame price. A music-API note dated from August 20, 2026 is not a Hailuo video term.\n\n### Prompt notes\n\nUse the host's example only as a shape. fal's Hailuo 2.3 example describes an electrical storm, camera position, and atmosphere in one paragraph. Higgsfield's H3 example is a one-line ocean wide shot. Neither was generated here. The operational limits that were printed: stay inside 7,000 characters on the Higgsfield H3 note, expect the optimizer to rewrite a 2.3 prompt when `prompt_optimizer` is left at its default, and do not assume a 9:16 request survives on image-to-video, because that schema says the keyframe wins.\n\nArena scores were not opened. Not observed.\n\n## Where to run\n\n1. Official price overview: [platform.minimax.io/docs/pricing/overview](https://platform.minimax.io/docs/pricing/overview). This is the H3 per-second list and the legacy clip list.\n2. fal Hailuo 2.3 standard text-to-video: [the fal model card](https://fal.ai/models/fal-ai/minimax/hailuo-2.3/standard/text-to-video). Legacy 768p clip prices that match the official 2.3 rows.\n3. fal Hailuo-02 standard text-to-video: [the fal model card](https://fal.ai/models/fal-ai/minimax/hailuo-02/standard/text-to-video). The $0.27 versus $0.28 disagreement lives here.\n4. Higgsfield's catalog: `POST https://api.higgsfield.ai/minimax/h3/text-to-video` and the image-to-video sibling, documented in the docs snapshot. The studio product is [/tools/higgsfield](/tools/higgsfield).\n5. Hailuo app: [/tools/hailuo](/tools/hailuo). That page holds the subscribe cards and the MiniMax API rates from the 2026-10-17 check. This model page's prices stay the 2026-10-16 official table.\n\nRunway's API guide, read 2026-10-16, prices `hailuo3` at 10 credits per second at 768P and 15 credits per second at 2K, and `h3_max` at 5 credits per second at 480p and 8 credits per second at 768p. Both `hailuo3` rows add 2 credits per reference image or keyframe, and they bill reference video at the same per-second rate as output, capped at 15 seconds. At $0.01 per credit the base rates derive to $0.10, $0.15, $0.05, and $0.08 per second. The $0.05 and $0.08 derived figures match official H3-Max 480P and the $0.08 H3 or H3-Max 768P rows. The derived $0.10 for `hailuo3` 768P does not match official H3 768P at $0.08, and the derived $0.15 for `hailuo3` 2K does not match official H3 2K at $0.13. Treat Runway's id as Runway's id. The guide is [Runway API pricing](https://docs.dev.runwayml.com/guides/pricing), and the product is [/tools/runway](/tools/runway).\n\nLuma lists Minimax H3 at 28 credits per second at 768p and 46 credits per second at 1440p (2K), audio column Yes, on [lumalabs.ai/pricing](https://lumalabs.ai/pricing). Those are Luma credits, not MiniMax dollars.\n\n## Multi-channel pricing\n\n**Official MiniMax, 2026-10-16 fetch.** The per-second and per-clip tables in Specs. A 10-second H3 clip at 768P is 10 times $0.08 = $0.80 derived. A 10-second H3 clip at 2K is $1.30 derived. A 6-second legacy Hailuo-2.3 at 768P is the printed $0.28, not a per-second invention. Input images beyond the free allowance add the printed per-image fees. Input audio adds $0 on the rows that say Free.\n\n**fal legacy cards, same day.** Hailuo 2.3 standard matches official $0.28 and $0.56. Hailuo-02 standard prints $0.045 per second and $0.27 per 6 seconds, which is 6 times 0.045. Official Hailuo-02 at 768P for 6 seconds is $0.28. Report the conflict. Do not \"correct\" fal to $0.28 or the official page to $0.27.\n\n**Runway credits.** See the derived figures above. They are a third channel only after you accept they are Runway's model ids times $0.01.\n\n**Higgsfield.** The H3 schema extract contains no credit or dollar field. Higgsfield's billing page says successful generations spend account credits, failed and `nsfw` requests are not charged, and outputs are kept at least seven days. An estimate example on that page uses Soul, an image model, and the page says the sample numbers illustrate the response format. They are not an H3 price. H3's Higgsfield dollar price is not observed.\n\nConsumer app plans were not opened on this 2026-10-16 pass. The yearly cards, and a 2026-10-17 read of the MiniMax API rates, are on [/tools/hailuo](/tools/hailuo). Do not paste those app cards onto the per-second rows above.\n\n## Workflow and platform fit\n\nSite labels only.\n\n**Faceless.** H3 on Higgsfield can be asked for 9:16, 5 to 15 seconds, at a fixed 2K on that schema. Official billing can be 768P at $0.08 per second or 2K at $0.13. The Higgsfield route's fixed 2K means you may not get the cheaper 768P official tier through that schema. Native audio is not established, so a faceless voice still comes from [/tools/elevenlabs](/tools/elevenlabs), and captions from [/tools/capcut](/tools/capcut).\n\n**UGC ads.** Image-to-video with an optional end frame is on the Higgsfield H3 schema. Five free input images on official H3, then $0.04, is enough for a small product board before the meter starts. A talking customer is still an avatar workflow. See [/models/heygen-video](/models/heygen-video) and [/tools/creatify](/tools/creatify).\n\n**AI avatar.** Not this model. Nothing in the price table describes a locked presenter.\n\n**Long-to-short.** Input video is billed per second of input at the output resolution. That is a restyle or continuation meter, not a highlight cutter. Cut first, then send a short input if you use the video-input price.\n\n**Captions and cover.** Not in the video schema. A cover is a separate image generation. MiniMax image-01 is priced, and it is a different model.\n\n**Voice and publish.** Because output audio was not described, the publish path is silent video plus a voice track. 9:16 exists on the Higgsfield text schema and is overridden by the keyframe on image-to-video. Shorts vertical fit is real for text-to-video on that host. Confirm the official API's own aspect list before you promise 9:16 on a direct MiniMax call. That official aspect list was not on the price page. Not observed there.\n\n**Shorts vertical fit.** The price is the friendly part. Eight cents per second at 768P is a low list price next to Veo standard and next to fal's Seedance 1080p figure. The missing pieces for a one-request Short are generated audio, an official duration cap, and an official aspect enum. You can still build a vertical Short by generating 5 to 15 seconds on the Higgsfield schema, adding voice and captions outside, and staying off 2K if the official 768P route is available to you.\n\n## Pros\n\n- The official page prints a per-second H3 price, so a 10-second estimate is multiplication rather than a bundled credit guess.\n- Legacy 2.3 prices on that page match fal's 2.3 standard card exactly, which is a useful cross-check.\n- Image inputs include a free allowance, so a few reference frames do not automatically become a second invoice line.\n- Failed-generation rules were not on the MiniMax extract. Higgsfield does state that `failed` and `nsfw` are not charged, for generations on that platform.\n- 9:16 is an explicit Higgsfield H3 text-to-video value, which many price pages omit.\n- Regeneration to 2K has its own lower per-second output price, so upscaling is not secretly billed as a full new 2K generation. Inputs are billed again. Read both lines.\n\n## Cons\n\n- Official duration and aspect are not on the price page. Anyone who quotes \"H3 makes 10 second 1080p clips\" is importing a legacy Hailuo number or a host schema. H3's official resolutions on the price page are 768P and 2K, not a row named 1080P. H3-Max adds 480P. Higgsfield's H3 schema is 2K only.\n- Native output audio was not documented. Luma's Audio Yes mark is not a substitute for that sentence.\n- fal Hailuo-02 at $0.27 for 6 seconds disagrees with the official $0.28. Pick the invoice you will actually receive.\n- Runway's `hailuo3` derives to $0.10 per second at 768P and $0.15 at 2K. Official H3 is $0.08 and $0.13 on those rows. Host tables drift.\n- The H3 fal path 404'd. You cannot cite a fal H3 schema from this pass.\n- Open weights were not verified. Do not brief Hailuo as a local model from this page.\n- Consumer plan allowances were not opened on this 2026-10-16 pass. The app cards are on [/tools/hailuo](/tools/hailuo), verified 2026-10-17. A subscriber's \"unlimited\" claim is not a row in the tables above.\n\n## Alternatives\n\n- [Kling](/models/kling) costs more per second on the official 1080p audio rows and includes a documented audio price. Choose Kling when speech in the same request is required.\n- [Seedance](/models/seedance) reaches 30 seconds with audio included in the fal US price. H3's Higgsfield cap in the schema read here is 15 seconds, and official duration is not observed.\n- [Veo](/models/veo) is the short, expensive, sound-on Google option with a GA region line.\n- [Luma Ray](/models/luma-ray) is the keyframe and EXR option, with no native audio either, and with a clearer modify story.\n- [Runway Gen](/models/runway-gen) is Runway's own model at a derived $0.12 per second base, separate from Runway's Hailuo credit rows.\n\n## Related tools\n\n- [/tools/hailuo](/tools/hailuo) is the Hailuo app. MiniMax API rates from the 2026-10-17 check live on that tool page. App yearly cards are not this model's per-second table.\n- [/tools/higgsfield](/tools/higgsfield) hosts the H3 text, image, and reference endpoints documented in this pass.\n- [/tools/runway](/tools/runway) prices `hailuo3` and `h3_max` in credits. Match the id before you compare to MiniMax's $0.08.\n- [/tools/capcut](/tools/capcut) is the caption and cover step.\n- [/tools/elevenlabs](/tools/elevenlabs) is the voice step while native audio stays not observed.\n- [/tools/creatify](/tools/creatify) and [/tools/heygen](/tools/heygen) cover structured ads and presenters.\n- [/tools/kling](/tools/kling) and [/tools/seedance](/tools/seedance) are the neighboring model apps when you leave the Hailuo price list.\n\n## FAQ\n\n### Is Hailuo 3 the same thing as MiniMax-H3?\n\nOn this page, yes. The official price rows use the id MiniMax-H3. Hosts sometimes say Hailuo 3 or `hailuo3`. Confirm the id on the invoice. Hailuo 2.3 and Hailuo-02 are older clip-priced models that remain on the same official page.\n\n### Does a second of video include sound?\n\nNot observed as generated audio. Input audio is free on the official H3 and H3-Max rows. That is a materials rule. Plan a separate voice track if you need speech you can direct.\n\n### Why does fal say $0.27 when MiniMax says $0.28?\n\nfal's Hailuo-02 standard card prints $0.045 per second and $0.27 for 6 seconds. The official Hailuo-02 768P 6-second row prints $0.28. Both were visible on 2026-10-16. Use the price of the host that will charge you. Hailuo 2.3 on fal matches the official $0.28 and $0.56.\n\n### Can I make a vertical clip?\n\nOn Higgsfield's H3 text-to-video schema, `9:16` is an allowed aspect. On that host's image-to-video schema, the keyframe's aspect wins. The official MiniMax price page does not list aspects. A direct-API vertical guarantee is not observed.\n\n### Is there a fal price for H3 itself?\n\nNot from this pass. The H3 fal `llms.txt` path returned 404. fal prices quoted here are Hailuo 2.3 and Hailuo-02. An older video-01-live card at $0.50 per video is not H3.", "path": "/models/minimax-hailuo" }, { "capabilities": [ "text-to-video", "video-generation" ], "verified_on": "2026-10-16", "native_audio": false, "multi_shot": false, "t2v": true, "i2v": false, "open_source": false, "in_launch_feed": true, "name": "Runway Gen", "summary": "Runway Gen-4.5 is Runway's own video model, billed at 12 credits per second on the API guide and as seconds of Gen-4.5 on consumer plans. A text-versus-image mode matrix for Gen-4.5 was not on that price card.", "example": false, "price_band": "API 12 credits per second at $0.01 per credit on 2026-10-16; consumer plans from $12 per month billed annually", "where_to_run": [ "Runway consumer plans", "Runway developer API" ], "slug": "runway-gen", "markdown": "## Overview\n\nRunway Gen-4.5 is the flagship video model sold by Runway, as distinct from the other companies' models that Runway's API also prices in the same credit. This page is about Gen-4.5. The product, the plan cards, and the editor are [/tools/runway](/tools/runway). Kling, Seedance, Veo, Wan, and Hailuo rows on the Runway price guide are not Gen. Follow those models to [/models/kling](/models/kling), [/models/seedance](/models/seedance), [/models/veo](/models/veo), and [/models/minimax-hailuo](/models/minimax-hailuo).\n\nStatus on 2026-10-16: Gen-4.5 is on the public consumer pricing page and on the developer pricing guide. That is general availability for people who can buy the listed plans or API credits. A waitlist was not on either page. A country block list was not on either page. Enterprise language on the consumer page says enterprises get two months of usage free if they sign by 10/31. That date is a sales window on the page, not a model sunset.\n\nFilter booleans are conservative where the card is silent. `t2v` is true because both surfaces sell Gen-4.5 as seconds of generated video, and the consumer page accounts for \"seconds of Gen-4.5\" inside a plan. `i2v` is false in the filter because the API guide bills `gen4.5` as one line, 12 credits per second, and does not publish a text-to-video row next to an image-to-video row. Image-to-video may exist in the product. It was not a mode on the price card, so the boolean does not assert it. `native_audio` is false. Models that have an audio split on that same guide, including `veo3.1` and `kling3.0_pro`, are written as two lines. Gen-4.5 is one line. A native-audio switch for Gen-4.5 was not observed. `multi_shot` is false. A multi-shot field for Gen-4.5 was not on the guide. Runway does sell a `multi_shot_video` recipe at a different credit rate. That recipe is not labeled Gen-4.5 in the cell that was read. `open_source` is false.\n\n## Specs\n\nVersion line. Gen-4.5 is the model this page prices. Gen-4 Turbo is a cheaper credit line on the same guide, 5 credits per second, and the consumer page also counts \"seconds of Gen-4 Turbo.\" Gen-4 image is a still model (`gen4_image`, 5 credits per 720p image or 8 credits per 1080p image). Aleph 2.0 shares the professional-output surcharge table with Gen-4.5 and is not Gen-4.5. Act Two is 5 credits per second and is not given a Gen-4.5 alias here.\n\nWhat the API guide states for Gen-4.5:\n\n| Field | Observed value |\n| --- | --- |\n| Price | 12 credits per second. |\n| Credit cash price | $0.01 per credit, so 12 credits is $0.12 per second derived. |\n| Professional formats | `prores` and `png_sequence` add 5 credits per second on Gen-4.5 and Aleph 2.0. |\n| HDR and 10-bit profiles | 20 credits per second, or 40 credits per second when the output is larger than 4 megapixels (roughly 4K). Profiles named: `hdr10`, `hlg`, `sdr_rec709_10bit`, `hdr_pq_12bit_master`, `hdr_prores`, `hdr_png_sequence`, `hdr_exr_sequence`, `hdr_exr_acescg_sequence_1_3`, `hdr_exr_acescg_sequence_2_0`. |\n\nDuration enum, aspect enum, frame rate, maximum prompt length, and a reference-image cap for Gen-4.5 were not on the pricing guide. They are not observed. The consumer page expresses plans as seconds of Gen-4.5, which confirms a per-second meter and does not reveal the maximum seconds per request.\n\nConsumer credit examples on [runwayml.com/pricing](https://runwayml.com/pricing), which redirected to runway.com/pricing on 2026-10-16: Gen-4.5 is 60 credits per 5 seconds, which is 12 credits per second and matches the API line. On the same compare table, Aleph 2.0 is 140 credits per 5 seconds, Seedance 2.0 Pro at 1080p is 160 credits per 4 seconds, and Seedance 2.0 Fast is 116 credits per 4 seconds. Those three rows are not Gen-4.5. Do not average them into the 12-credit rate.\n\nA calculator embedded in the API guide uses `'gen4.5': 12`, which agrees with the table.\n\n### Prompt notes\n\nNo Gen-4.5 prompt was published as a worked example in the pricing extracts, and none was run here. The practical note is about the meter. A longer prompt does not show a token price. You pay credits per second of output, plus a format surcharge if you ask for ProRes, a PNG sequence, or an HDR profile. Asking for \"4K HDR\" without reading the surcharge is how a 12-credit estimate becomes 12 plus 40. The guide says the 40-credit HDR rate applies when the output is larger than 4 megapixels. Whether every Gen-4.5 render can reach that size was not stated as a resolution enum.\n\nArena scores for Gen-4.5 were not opened. Not observed.\n\n## Where to run\n\n1. Runway's consumer pricing page: [runwayml.com/pricing](https://runwayml.com/pricing). Plans, included seconds, and the Gen-4.5 credit examples. The app checkout links point at `app.runwayml.com`.\n2. Runway's developer pricing guide: [docs.dev.runwayml.com/guides/pricing](https://docs.dev.runwayml.com/guides/pricing). The `gen4.5` row, the $0.01 credit price, and the professional-output surcharges.\n3. The product write-up in this directory: [/tools/runway](/tools/runway).\n\nA second company must appear because a selection page should not trap the reader inside one vendor. Gen-4.5 itself was not found as a resale row on fal, on BytePlus, or on the Higgsfield catalog index saved the same day. For Gen-4.5, the two Runway surfaces are the channels, and they are different wallets: plan credits versus API credits purchased at $0.01. That is \"official only\" across two Runway doors, not two unrelated hosts. If a workflow can change models, the other doors on the same API guide are the escape hatch, and they are different models.\n\nLuma's pricing page did not show a Gen-4.5 credit row in the portion read. Not observed there.\n\n## Multi-channel pricing\n\nVerified 2026-10-16 on both Runway pages. There is no non-Runway dollar card for Gen-4.5 in this pass. The two channels are the API meter and the consumer plan.\n\n**API.** $0.01 per credit. Sales tax may apply. `gen4.5` is 12 credits per second. Derived base price $0.12 per second. Add 5 credits per second for `prores` or `png_sequence`. Add 20 credits per second for the named 10-bit and HDR profiles, or 40 credits per second when output is larger than roughly 4K. A 5-second base clip is 60 credits, $0.60 derived, which matches the consumer page's \"60 credits/5s\" Gen-4.5 row. The \"140 credits/5s\" row on that table is Aleph 2.0, not a second Gen-4.5 rate. A 5-second clip with the 40-credit HDR adder would be (12+40) times 5 = 260 credits, $2.60 derived. That stack is arithmetic from the guide, not a SKU name Runway printed.\n\n**Consumer plans, from the pricing page text.** Free: $0 per month, 125 one-time credits. Standard: $15 per month, or $12 per month billed annually (save $36 a year), 625 credits monthly. The page equates 625 credits to 52 seconds of Gen-4.5, 104 seconds of Gen-4 Turbo, or 78 Gen-4 images at 1080p. Pro, labeled Pro on the card: $35 per month, or $28 per month billed annually (save $84 a year), 2250 credits monthly, equated to 187 seconds of Gen-4.5, 375 seconds of Gen-4 Turbo, or 281 Gen-4 images at 1080p. The Pro card also lists top-up credits, unlimited 4K upscaling, and a bullet whose visible text is \"Unlimited\" followed by chips for Seedance 2.5, Gen-4.5, Kling 3.0, Nano Banana Pro, and more. Max: $95 per month, or $76 per month billed annually (save $228 a year), 9500 credits, equated to 791 seconds of Gen-4.5, 791 seconds of Gen-4, 1900 seconds of Gen-4 Turbo, 1187 Gen-4 images at 1080p, or 475 Nano Banana Pro images at 2K. Max rolls unused credits for one month. The page says unlimited Wan 3.0 and Wan 3.0 Prime in Tool Mode until October 18. Yearly credit totals printed nearby: $12 per month maps to 7,500 credits a year, $28 per month to 27,000, and $76 per month to 114,000.\n\nCheck the arithmetic the page invites. 625 credits divided by 12 credits per second is about 52.1 seconds, which matches the \"52s of Gen-4.5\" line. 2250 / 12 is 187.5, matching 187 seconds. 9500 / 12 is about 791.7, matching 791 seconds. Those plan equivalences use the 12-credit Gen-4.5 rate. The 140-credit row is Aleph 2.0. The FAQ on the page says Standard and Pro monthly credits do not roll over, and that extra credits can be bought with a 1,000 credit minimum. Upgrade is immediate and prorated, per that FAQ.\n\nThe Pro \"Unlimited\" chip row names Seedance 2.5, Gen-4.5, Kling 3.0, and Nano Banana Pro together. The page does not define a fair-use cap in the extracted lines. The same Pro card still equates 2250 credits to 187 seconds of Gen-4.5, so the chip row and the seconds line disagree. Seedance's own API credit price is on the Seedance model page.\n\n## Workflow and platform fit\n\nSite workflow labels only. No workflow guide URLs are linked, because those guides are not published as real pages.\n\n**Faceless.** A plan that buys seconds of Gen-4.5 can supply B-roll, but native audio was not on the Gen-4.5 price line. Budget a voice. [/tools/elevenlabs](/tools/elevenlabs) is the voice product in this directory. Captions and a cover are [/tools/capcut](/tools/capcut). A 52-second Standard monthly pool is one short faceless test, not a posting cadence. Pro's 187 seconds is still a handful of cuts if you iterate.\n\n**UGC ads.** Runway's API guide prices recipes that are not Gen-4.5: `product_ad`, `product_swap`, and `product_ugc`, each with a 4-second minimum and a 15-second maximum, at a few hundred credits for the first 4 seconds plus a per-extra-second rate. Read those recipes on the guide before you assume a UGC job bills at 12 credits per second. [/tools/creatify](/tools/creatify) is the template path when you do not want to assemble the recipe yourself. Image-to-video for Gen-4.5 specifically remains not observed as a separate row.\n\n**AI avatar.** Gen-4.5's card does not describe a photo avatar or a voice clone as part of the model meter. The consumer Pro plan mentions one custom voice clone as a plan feature. That is a product entitlement, not a Gen-4.5 spec. Presenters belong on [/models/heygen-video](/models/heygen-video).\n\n**Long-to-short.** Nothing on the Gen-4.5 line ingests a long recording. Use an editor to cut, then generate replacement B-roll by the second.\n\n**Captions and cover.** Gen-4 image can make a still at 5 or 8 credits. It is not an auto-caption model. Captions stay in CapCut or a caption tool. Not observed inside Gen-4.5.\n\n**Voice and publish.** Because native audio is not on the price split, publish flow is picture first and sound second. Vertical aspect support for Gen-4.5 was not on the price guide. Shorts fitness is therefore \"possible if the product UI offers 9:16,\" and that UI control was not copied into the pricing extract. Do not claim 9:16 as a verified enum. HDR and ProRes matter for a grade and are usually stripped or recompressed before a Shorts upload. The surcharge is a master cost, not a phone-preview cost.\n\n**Shorts vertical fit.** Economical at the derived $0.12 per second base if the unobserved aspect control exists in the app you use, and incomplete as a one-stop Shorts tool because audio, captions, and a confirmed 9:16 enum were not on the card. Compare Kling or Seedance when the vertical enum and the audio flag have to be in the schema you cite.\n\n## Pros\n\n- The API rate is a single number, 12 credits per second, and the consumer \"seconds of Gen-4.5\" lines divide cleanly by that number. Two Runway pages agree on the base meter.\n- The $0.01 credit price makes the derived dollar predictable, before tax and before surcharges.\n- Professional and HDR outputs are priced as adders with named format strings, which is the right shape for a finishing pipeline.\n- Plan equivalences are printed in seconds, so a buyer can see that Standard's 625 credits are 52 seconds of this model and not a vague \"credit pack.\"\n- The same guide prices other models, which makes a controlled comparison inside one wallet possible. Those other models still need their own pages, because a credit is not a quality claim.\n\n## Cons\n\n- The consumer compare table prints 140 credits per 5 seconds on the Aleph 2.0 row, next to Gen-4.5 at 60 credits per 5 seconds. Quote the model name with the row. There is not a second unmapped Gen-4.5 rate on that table.\n- Duration, resolution, and aspect for a default Gen-4.5 request were not on the pricing guide. The spec table is thinner than Veo's or Seedance's.\n- Image-to-video is not a separate price line. The filter refuses to call it true.\n- Native audio is not a price split. A talking clip needs another tool or a model whose card itemizes audio.\n- Gen-4.5 was not observed on a non-Runway host. You cannot arbitrage this model across fal and Runway the way you can arbitrage Kling.\n- The Pro \"Unlimited\" chips include Gen-4.5 by name, next to Seedance 2.5. The Gen-4.5 meter on that same card is still 187 seconds at 2250 credits. Read both lines before you treat Gen-4.5 as unlimited.\n\n## Alternatives\n\n- [Kling](/models/kling) on the same Runway guide is 9 to 43 credits per second depending on standard, pro, audio, and 4K. Kling's own site also prints unit prices. Choose Kling when you need a documented audio toggle and a 9:16 fal enum.\n- [Seedance](/models/seedance) on Runway is 20, 30, or 68 credits per output second by resolution, higher than Gen-4.5's base 12, with a longer duration enum on fal. The Pro plan's unlimited Seedance line, if it applies to your account, changes that comparison inside the app only.\n- [Veo](/models/veo) on Runway is 20 or 40 credits per second for standard silent versus audio, above Gen-4.5's base rate, with a Google GA card elsewhere.\n- [Luma Ray](/models/luma-ray) is the HDR and EXR specialist with its own dollar API. Gen-4.5's HDR path is a credit surcharge on Runway, not an EXR-first product page.\n- [MiniMax Hailuo](/models/minimax-hailuo) is the lower official list price per second when you do not need Runway's format menu.\n\n## Related tools\n\n- [/tools/runway](/tools/runway) is the product that contains Gen-4.5, the plans, and the other hosted models.\n- [/tools/capcut](/tools/capcut) finishes captions, covers, and vertical reframes after export.\n- [/tools/elevenlabs](/tools/elevenlabs) supplies voice because Gen-4.5's price line has no audio split.\n- [/tools/creatify](/tools/creatify) is the UGC-template alternative when the Runway product recipes are more structure than you want to bill by hand.\n- [/tools/heygen](/tools/heygen) is the avatar product, not a Gen-4.5 mode.\n- [/tools/kling](/tools/kling), [/tools/seedance](/tools/seedance), and [/tools/higgsfield](/tools/higgsfield) are the neighboring studios. Higgsfield was not observed reselling Gen-4.5 in the catalog snapshot.\n\n## FAQ\n\n### Is Gen-4.5 the same as every model inside Runway?\n\nNo. Gen-4.5 is one row. Kling, Veo, Seedance, Wan, and Hailuo have their own rows on the same guide. The app can hide that distinction behind a single credit balance.\n\n### How much is one second?\n\nTwelve credits on the API table. At the printed $0.01 per credit, that is $0.12 before tax and before ProRes or HDR adders. The consumer compare table's 140 credits per 5 seconds row is Aleph 2.0. Do not average it with the Gen-4.5 rate of 12 credits per second.\n\n### Does it include native audio?\n\nNot observed as a Gen-4.5 price split. The guide does split audio for other model ids. This page sets native audio to false so the filter does not promise speech.\n\n### Can I start from an image?\n\nNot observed as a distinct Gen-4.5 price line. The boolean for image-to-video is false for that reason. Some other Runway rows do mention an image-to-video start frame. They are not `gen4.5`.\n\n### Do monthly credits roll over?\n\nThe consumer FAQ says Standard and Pro monthly credits do not roll over. Max can roll unused credits for one month. API credits bought at $0.01 are a different wallet. Their expiry was not in the sentences extracted here.", "path": "/models/runway-gen" }, { "capabilities": [ "text-to-video", "image-to-video", "native-audio" ], "verified_on": "2026-10-16", "native_audio": true, "multi_shot": true, "t2v": true, "i2v": true, "open_source": false, "in_launch_feed": true, "name": "Seedance", "summary": "ByteDance Seedance is a cinematic text-to-video and image-to-video family with synchronized audio on the 2.5 cards opened on 2026-10-16. Compare fal, Runway credits, and BytePlus before you pick a wallet.", "example": false, "price_band": "fal US Seedance 2.5 about $0.26 to $1.40 per second by resolution on 2026-10-16; Runway bills 20 to 68 credits per output second", "where_to_run": [ "fal Seedance 2.5 US text-to-video", "BytePlus ModelArk", "Runway API credits", "Higgsfield API Seedance 2.5" ], "slug": "seedance", "markdown": "## Overview\n\nSeedance is ByteDance's commercial video-generation family, sold under Dreamina and ModelArk model ids and resold by API hosts. This page is a selection sheet for the model. It is not a generation console, and it is not a review of the Dreamina or CapCut product. The product surface lives at [/tools/seedance](/tools/seedance). Use that page for the app. Use this page when you need capability, where a job can run, and what each channel charged on a named date.\n\nCapability tags on the filter are text-to-video, image-to-video, and native audio. `t2v` and `i2v` are both true because opened schemas name both jobs for Seedance 2.5. `native_audio` is true because the fal 2.5 US text-to-video card and the Higgsfield 2.5 image-to-video schema both expose `generate_audio`, and fal states that the video price does not change when audio is on. `multi_shot` is true for the family because the fal Seedance 2.0 text-to-video card, fetched the same day, describes \"native audio, multi-shot editing.\" The Seedance 2.5 US text-to-video schema opened on fal does not include a `multi_prompt` field. Treat multi-shot as a 2.0 claim, and treat a 2.5 multi-shot parameter as not observed on that US card. `open_source` is false. No weight release was on the pages opened for this pass.\n\nStatus, as observed on 2026-10-16: generally available on fal's US Seedance 2.5 text-to-video endpoint and listed on BytePlus ModelArk pricing (page footer last updated October 8, 2026 13:17:02) under model id `dreamina-seedance-2-5-260628`. A region block list was not on the pages saved here. Do not read a missing region table as a US ban or as worldwide availability. Higgsfield's public API catalog, saved the same day from `llms-full.txt`, documents Seedance 2.5 text-to-video, image-to-video, reference-to-video, video edit, and video extend. That is a host catalog, not ByteDance's own region policy.\n\n## Specs\n\nVersion line. The current paid generation discussed here is Seedance 2.5. Seedance 2.0 remains on fal and on the Runway credit table, with a shorter duration cap and a 4K tier that 2.5's fal US card does not list. BytePlus also prints ids `dreamina-seedance-2-0-260128`, `dreamina-seedance-2-0-fast-260128`, and `dreamina-seedance-2-0-mini-260615`. Fast and Mini dollar-per-second rows were not cleanly tied to resolutions in the ModelArk HTML zones parsed for this page, so those variant prices are not observed here.\n\nSeedance 2.5 on fal, US text-to-video, model id `bytedance/seedance-2.5/us/text-to-video`, endpoint `https://fal.run/bytedance/seedance-2.5/us/text-to-video`:\n\n| Field | Observed value |\n| --- | --- |\n| Duration | 4 to 30 seconds, or `auto`. Default `auto`. |\n| Resolution | `480p`, `720p`, `1080p`. Default `720p`. |\n| Aspect ratio | `auto`, `21:9`, `16:9`, `4:3`, `1:1`, `3:4`, `9:16`. Default `auto`. |\n| Audio | `generate_audio` default true. Cost is the same with or without audio, per fal's own sentence. |\n| Bitrate | `standard` or `high`. Default `standard`. |\n| Codec | `auto`, `H264`, or `H265`. Default `auto`. |\n| Token formula | `(height * width * duration * 24) / 1024`. |\n\n4K, frames per second as a user field, and a reference-image array were not fields on this US text-to-video schema. Frames per second inside the token formula is 24. That is a billing constant on the card, not a published output-fps guarantee beyond the formula.\n\nSeedance 2.5 image-to-video on the Higgsfield API catalog (`POST https://api.higgsfield.ai/bytedance/seedance-2.5/image-to-video`): prompt optional; framing follows `image_url`; `end_image_url` is an optional last frame; duration integer minimum 4 and maximum 30; resolution `480p`, `720p`, or `1080p`; bitrate `standard` or `high`; `generate_audio` is a boolean described as \"Generate audio with the video.\" A dollar price is not on that schema page.\n\nSeedance 2.0 on fal text-to-video: duration 4 to 15 seconds or `auto`; resolution adds `4k`; the card text says native audio and multi-shot editing; token formula matches 2.5. A separate 2.0 image-to-video schema was not the file saved from fal. Higgsfield's catalog does list a Seedance 2.0 image-to-video page. Its parameter table was not copied into this selection sheet.\n\nReference limits, subject locks, and a maximum number of input images for the ByteDance-hosted API were not observed on the fal or ModelArk pages used here. Higgsfield's Cinema Studio page points editors at \"separate Seedance 2.5 endpoints\" for editing and extension, and its Seedance 2.5 reference-to-video page exists in the catalog. The reference-count cap for that endpoint was not extracted into this sheet. Call it not observed rather than copying a third-party \"50 references\" line.\n\n### Prompt notes\n\nfal's published 2.5 example prompt is a vendor sample, not a result this directory generated: \"An octopus finds a football in the ocean and excitedly calls its octopus friends to come and play. Cut scene to an octopus football game under the sea.\" The useful check is structural. The sample asks for a cut. Whether 2.5 honors a cut inside one request is not measured here. The 2.0 card is the one that uses the words \"multi-shot editing.\"\n\nArena or Artificial Analysis scores for Seedance were not opened on 2026-10-16. Do not treat a missing score as a rank.\n\n## Where to run\n\nYou need at least two real entry points. These were opened or saved on 2026-10-16.\n\n1. fal's US text-to-video card: [Seedance 2.5 US text-to-video](https://fal.ai/models/bytedance/seedance-2.5/us/text-to-video). This is the schema and the per-second rough prices quoted below. An older fal path, `fal-ai/bytedance/seedance/v2.5/text-to-video`, returned HTTP 404. Use the `bytedance/seedance-2.5/us` id.\n2. BytePlus ModelArk pricing: [ModelArk model pricing](https://docs.byteplus.com/en/docs/modelark/model-pricing). Footer last updated October 8, 2026. This is the official token and example-clip surface for `dreamina-seedance-2-5-260628`.\n3. Runway's developer price guide: [Runway API pricing](https://docs.dev.runwayml.com/guides/pricing). Seedance 2.5 is a credit row, not Runway's own Gen model. The app that spends those credits is [/tools/runway](/tools/runway).\n4. Higgsfield's model catalog documents Seedance 2.5 image-to-video at `https://api.higgsfield.ai/bytedance/seedance-2.5/image-to-video` inside the `llms-full.txt` snapshot saved on 2026-10-16. The studio product is [/tools/higgsfield](/tools/higgsfield). Hosting Seedance there does not make Seedance a Higgsfield-owned model.\n\nDreamina's consumer subscription page was not re-fetched for this model pass. Membership dollars that appear on the tool page are not repeated here as freshly verified. CapCut's editor is a different product: [/tools/capcut](/tools/capcut). A CapCut export preset that embeds Seedance was not observed in the files for this page.\n\n## Multi-channel pricing\n\nPrices below were read on 2026-10-16 unless a source footer names an earlier update. Channels are separate wallets. A fal second is not a Runway credit and not a Dreamina subscription second.\n\n**fal, US Seedance 2.5 text-to-video.** fal writes: roughly **$0.5676 per second** at 720p, **$0.2646 per second** at 480p, and roughly **1.396278 per second** at 1080p. The 1080p figure is quoted as printed. fal omitted a currency sign on that one number. Token rates on the same card: **$0.02568 per 1000 tokens** at 480p and 720p, and **$0.02808 per 1000 tokens** at 1080p. A 10-second 720p clip at the printed per-second figure is about $5.68 if you multiply 10 by 0.5676. That product is derived. The card says \"roughly,\" so the multiplication inherits the hedge. Audio does not change this price, per the schema text.\n\n**fal, Seedance 2.0 text-to-video.** **$0.3034 per second** at 720p and **$0.682 per second** at 1080p. **$0.014 per 1000 tokens** for 480p, 720p, and 1080p. **$0.008 per 1000 tokens** for 4K. A 480p per-second dollar figure was not in that paragraph. Do not back-solve one from the token rate unless you also fix width and height, which the card does not pin to a single pixel size.\n\n**Runway API, verified on the same pricing guide.** Credits cost **$0.01 per credit**. Seedance 2.5 output rates, with resolution labels in the table:\n\n| Output | Credits per output second | Extra credits per second of input and reference video | Minimum |\n| --- | --- | --- | --- |\n| 480p | 20 | 10 | 80 credits |\n| 720p | 30 | 15 | 80 credits |\n| 1080p | 68 | 34 | 80 credits |\n\nReference images and audio are free on those rows. Combined input video is capped at 30 seconds in the cell text. Derived output-only dollars at $0.01 per credit: $0.20, $0.30, and $0.68 per second. Label them derived. They are not fal's prices and they ignore the input-video adder. Seedance 2.0 on the same table: 36 credits per second at 480p/720p, 40 at 1080p, 150 at 4K. Fast 480p/720p is 29 credits per second. Mini 480p/720p is 16 credits per second with a 64 credit minimum. Those Fast and Mini rows are Runway's credit card, not the BytePlus zones that stayed unmapped.\n\n**BytePlus ModelArk.** For `dreamina-seedance-2-5-260628` the document says pricing varies by output resolution and by whether the input includes video. The inserts read: for 480p and 720p outputs, input without video 10.70 and input with video 6.40; for 1080p outputs, input without video 11.7 and input with video 7.0. The unit label was not in that editor zone. Other sections of the same site bill language models in USD per million tokens. This page does not promote 10.70 into a dollar-per-second rate and does not assert the million-token unit, because the header was not in the zone. A sibling cell reads \"Not supported yet.\" It was not tied to an \"offline\" column header in the parse, so offline support is not observed rather than denied.\n\nSeparate example inserts on the same HTML file read **0.514 per video** and **0.103 per second**, **1.156 per video** and **0.231 per second**, and **2.843 per video** and **0.569 per second**. Resolution labels for those three pairs sit in other zones. They are quoted as printed and not assigned to 480p, 720p, or 1080p here.\n\nDreamina consumer cards and any \"$0.04 per second\" banner were not re-opened. They are not observed for this pass. Luma's own pricing page also lists Seedance 2.5 in Luma credits (36, 78, and 191 credits per second at draft/480p, 720p, and 1080p for text-to-video and image-to-video, audio column Yes). That is a fifth wallet, in credits, verified on [lumalabs.ai/pricing](https://lumalabs.ai/pricing) on 2026-10-16. It is not ByteDance's price. A Luma-credit to USD conversion for Seedance was not printed on that card.\n\n## Workflow and platform fit\n\nSite workflow labels below are planning names. Real workflow guides for those slugs are not published, so this section does not link them.\n\n**Faceless.** Seedance 2.5 can hold a 9:16 frame and a clip up to 30 seconds with audio on the fal US card. That fits a faceless B-roll beat or a short scene with generated sound. It does not replace a voice track, captions, or a cover. Pair the clip with [/tools/elevenlabs](/tools/elevenlabs) when you need a controlled voice, and with [/tools/capcut](/tools/capcut) when you need captions and a cover. Whether Seedance's generated speech is clean enough to publish alone was not measured here.\n\n**UGC ads.** Image-to-video from a product still is the closer job, and Higgsfield's 2.5 schema supports a start frame plus an optional end frame, 4 to 30 seconds, including 9:16 on the text card's aspect list. The image-to-video schema's aspect field was not in the short extract used here; the text-to-video card is the one that lists `9:16`. A talking founder is a different model. See [/models/heygen-video](/models/heygen-video).\n\n**AI avatar.** Seedance is a world and shot model, not an avatar engine. Lip-sync is a fal tag on the 2.5 US card (\"lipsync\" appears in the tag list) but a face-consistency guarantee was not in the schema. For a hosted presenter, start with HeyGen. For a performed clip you already shot, Luma Ray's Modify path is the closer comparison.\n\n**Long-to-short.** Seedance does not ingest a long podcast and cut highlights. Max observed generation length on 2.5 is 30 seconds. Use it to rebuild a beat, then cut in CapCut.\n\n**Captions and cover.** No caption or cover tool was on the video schemas. Those stay in the editor.\n\n**Voice and publish.** Native audio can ship inside the mp4 when `generate_audio` is true. fal says Chinese is not a separate documented language field on the US card that was read. A language list for Seedance speech was not observed. Publish vertical video when you set 9:16. Shorts, Reels, and TikTok are the same frame on this card. Safe-area margins for interface chrome were not observed.\n\n**Shorts vertical fit.** Strong on paper: 9:16 is an enum value, duration covers a Shorts hook through a 30-second spot, and 1080p is available. The cost jump is the practical limit. fal's printed 1080p figure is much higher than 720p. Many Shorts workflows can stay at 720p and spend the difference on another take.\n\n## Pros\n\n- Two opened 2.5 schemas agree on a long clip: 4 to 30 seconds, which is longer than Veo's 4/6/8 second card and longer than Kling 3 Pro's 3 to 15 second fal enum.\n- 9:16 is a first-class aspect on the fal US text-to-video card, so a vertical Shorts brief does not depend on a later crop. Cropping still changes composition. The enum only proves the model can be asked for that frame.\n- Audio is priced the same as silent video on the fal 2.5 card, which makes an A/B of \"with sound\" cheaper to attempt than on Veo, where Gemini's listed rates are with-audio prices and fal charges a lower silent rate.\n- You can buy the same family through fal dollars, Runway credits, BytePlus, and a Higgsfield endpoint. That is the point of a multi-channel sheet: when one host queues or reprices, another route is documented.\n- Image-to-video with an optional last frame is a real endpoint in the Higgsfield catalog, so a start/end board is not only a marketing sentence.\n\nThese points are checks against opened schemas and price cards. They are not a quality leaderboard.\n\n## Cons\n\n- The 1080p fal figure is high enough that a 10-second master is a different budget from a 720p draft. Teams that copy a 720p quote into a 1080p job will miss by a wide margin. Use the card, not a blended \"about fifty cents.\"\n- Multi-shot is documented on the 2.0 fal blurb and not as a field on the 2.5 US text-to-video schema. If your edit depends on per-shot prompts, Kling's `multi_prompt` field is the one that was actually in a 2026-10-16 schema. Seedance 2.5 may still cut inside a single prompt. That behavior was not measured.\n- BytePlus prints resolution-linked numbers whose unit was not in the same zone, plus per-video examples whose resolution was not in the same bullet. You cannot audit an invoice from those inserts alone.\n- Dreamina consumer allowances were not re-verified in this pass. A subscription second and an API second are different products even when both say Seedance.\n- Open weights were not observed. You cannot pin a local build to this model id from the pages read here.\n- fal's 2.0 and 2.5 token prices differ, and older articles that quote a single per-second number will not match the live `llms.txt`. When a blog and the live card disagree, the live card is the one quoted above.\n\n## Alternatives\n\n- [Kling](/models/kling). Kling 3 Pro on fal is a per-second price with an audio upcharge and an explicit multi-shot prompt list up to 15 seconds. Seedance 2.5 is the longer single clip and, on fal, does not charge extra for audio.\n- [Veo](/models/veo). Veo 3.1 on Vertex is 4, 6, or 8 seconds with a GA card in `us-central1` and a higher with-audio second on the Gemini price list. Pick Veo when the Google contract and SynthID path matter more than a 30-second take.\n- [Runway Gen](/models/runway-gen). Gen-4.5 is Runway's own model at 12 credits per second, about $0.12 derived. Seedance on Runway is a different row and a different model. Do not substitute the names because they share a credit wallet.\n- [Luma Ray](/models/luma-ray). Ray3.2's own FAQ says text-to-video and image-to-video have no native audio. Seedance is the closer pick when the clip must leave the API with generated sound. Ray is the closer pick when you need 16 keyframes, HDR, or EXR.\n- [MiniMax Hailuo](/models/minimax-hailuo). Hailuo 3 on the official pay-as-you-go page is a lower per-second list price at 768P and 2K, with duration not stated on that price table. Seedance is the one with a 30-second enum on the cards opened here.\n\nComparison URLs such as `/models/seedance-vs-kling` are not published. Differences above are the selection notes.\n\n## Related tools\n\n- [/tools/seedance](/tools/seedance) is the Dreamina / Seedance product. Read it for the app workflow. Do not copy subscription dollars from it onto this page unless you re-open the pricing URL.\n- [/tools/higgsfield](/tools/higgsfield) hosts a Seedance 2.5 endpoint beside Higgsfield's own Cinema Studio and Genjutsu models. The host is not the model.\n- [/tools/runway](/tools/runway) can spend credits on Seedance 2.5. Confirm the model id in the API call. The Gen-4.5 row is a different model.\n- [/tools/capcut](/tools/capcut) is the usual editor for captions, covers, and a vertical finish after the clip exists.\n- [/tools/elevenlabs](/tools/elevenlabs) is the separate voice path when generated Seedance audio is the wrong speaker or the wrong language.\n- [/tools/kling](/tools/kling) and [/tools/creatify](/tools/creatify) are adjacent products for a performance-style or ad-template finish. They do not run Seedance unless their own docs say so. That cross-run was not observed.\n\n## FAQ\n\n### Is Seedance 2.5 the same product as the Dreamina subscription?\n\nNo. Dreamina is a consumer wallet. fal, Runway, and BytePlus are different wallets. This page verifies fal's US 2.5 card, Runway's credit rows, and the ModelArk inserts named above. Dreamina plan dollars were not re-fetched on 2026-10-16.\n\n### Does audio cost extra?\n\nOn the fal US 2.5 text-to-video card, fal says the cost is the same whether audio is generated or not. Runway's Seedance 2.5 cells say reference audio is free and bill output seconds plus input video seconds. A BytePlus audio upcharge was not in the inserts quoted here.\n\n### Can I generate 9:16 for Shorts?\n\nThe fal US text-to-video schema lists `9:16` as an aspect option. Higgsfield's image-to-video extract used here did not restate the aspect enum. Set the aspect on the endpoint you actually call, and keep a 720p test before you pay the 1080p rate.\n\n### Which multi-shot claim is safe to repeat?\n\nSafe: fal's Seedance 2.0 text-to-video overview says \"multi-shot editing.\" Not safe: claiming the 2.5 US text-to-video schema exposes per-shot prompts. That field was not on the schema. Kling's fal 3 Pro schema does expose `multi_prompt`.\n\n### Where do I see quality scores?\n\nArtificial Analysis and crowd arena numbers were not opened for Seedance on 2026-10-16. They are not observed. A score without a date would go stale faster than the price card, which is already channel-specific.", "path": "/models/seedance" }, { "capabilities": [ "text-to-video", "image-to-video", "native-audio" ], "verified_on": "2026-10-16", "native_audio": true, "multi_shot": false, "t2v": true, "i2v": true, "open_source": false, "in_launch_feed": true, "name": "Veo", "summary": "Veo 3.1 is Google's video model with text-to-video, image-to-video, and sound generation. Vertex lists generate-001 as GA in us-central1; the Gemini API price page lists preview ids with no free tier.", "example": false, "price_band": "Gemini API paid with audio from $0.05 to $0.60 per second by tier, page updated 2026-10-15; fal standard $0.20 or $0.40 per second", "where_to_run": [ "Vertex Veo 3.1 generate-001", "Gemini API paid tier", "fal Veo 3.1", "Runway API Veo credit rows", "Google Flow app" ], "slug": "veo", "markdown": "## Overview\n\nVeo is Google DeepMind's video generation line. This selection page covers Veo 3.1 as it was documented on 2026-10-16. It is not a Flow or Gemini playground, and it does not sell a try-on credit. Google Flow is the app for Veo 3.1: [/tools/google-flow](/tools/google-flow). Plan stickers are on that tool page, verified 2026-10-17. People also reach Veo through Gemini, Vertex, or an API host. Those doors are listed below with the prices that were on a page. This model page does not copy Flow credits into a per-second rate.\n\nCapability tags: text-to-video and image-to-video are true. The Vertex model card for `veo-3.1-generate-001` lists both, plus first-and-last-frame, as supported video generation. Native audio is true because that same card lists \"Sound generation: Supported,\" and the Gemini API price table bills \"video with audio\" as the default paid rate. The card's modality row says Audio is not supported. Read that as input audio, next to Text and Image marked input-only and Video marked output-only. It does not cancel the sound-generation capability. `multi_shot` is false. Extend is a supported operation, and extend is not a one-request multi-shot prompt list. The fal 3.1 schema has duration, resolution, and aspect, and it has no `multi_prompt` field. `open_source` is false.\n\nStatus is split, and the split is the point of the page. On the Gemini Enterprise Agent Platform card fetched 2026-10-16, `veo-3.1-generate-001` is launch stage GA, release date November 17, 2025, retirement date November 17, 2026 or later. The footnote says that retirement is support on that platform and that the model may remain on the Gemini API afterward. Region on that card: United States, `us-central1`. Pay-as-you-go is \"Not supported\" on that cloud card. Provisioned throughput and fixed quota are supported. The Gemini API price page, last updated 2026-10-15 UTC, lists `veo-3.1-generate-preview`, `veo-3.1-fast-generate-preview`, and `veo-3.1-lite-generate-preview`, and it marks the free tier \"Not available.\" Preview ids and a GA id are both real. They are not the same SKU. A consumer waitlist was not on these two pages. DeepMind's model page is a third surface, titled Veo 3.1, with entry points to try it in Gemini, try it in Google Flow, and build with Veo. Parts of that decoded page still say \"Introducing Veo 3.\" Treat leftover 3.0 headings as page residue, not as the 3.1 spec.\n\n## Specs\n\nVersion line. Quote `veo-3.1-generate-001` when you mean the GA Vertex model. Quote the preview ids when you mean the Gemini API price rows. Fast and Lite are cheaper rows on the Gemini page. On the [Gemini Enterprise Agent Platform Veo 3.1 card](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/veo/3-1-generate), fetched 2026-10-16, `veo-3.1-generate-001` and `veo-3.1-fast-generate-001` are both launch stage GA, release date November 17, 2025. Fast's output resolutions on that card are 720p and 1080p. The 4K output line in the table below is the `veo-3.1-generate-001` card. `veo-3.1-lite-generate-001` is the card marked Preview. Gemini API price rows stay on the preview ids named below. Do not merge those ids.\n\nVertex technical specifications for `veo-3.1-generate-001`:\n\n| Field | Observed value |\n| --- | --- |\n| Length | 4, 6, or 8 seconds. Reference image to video only supports 8 seconds. |\n| Outputs per prompt | Maximum 4. |\n| Input image | Maximum 20 MB. |\n| Aspect | 9:16 and 16:9. |\n| Input resolution | 720p, 1080p. |\n| Output resolution | 720p, 1080p, 4K. |\n| Frame rate | 24 FPS. |\n| Container | `video/mp4`. |\n| Prompt language | English. |\n| Capabilities | Text to video, image to video, first and last frame, extend, reference images (asset images), sound generation, Content Credentials (C2PA). |\n| Region | `us-central1`. |\n| Quota line | \"50 tokens per minute\" on the regional online-prediction quota. This page does not translate that line into a video-second cap. |\n\nPay-as-you-go dollars are not on the Vertex card. The Gemini API table is the Google dollar list, and it is attached to preview ids.\n\nfal model `fal-ai/veo3.1`, read from its `llms.txt` on 2026-10-16: duration `4s`, `6s`, or `8s`, default `8s`. Resolution `720p`, `1080p`, or `4k`, default `720p`. Aspect `16:9` or `9:16`, default `16:9`. That duration set matches Vertex. A fast-tier fal schema was not fetched. Do not import a fast price from a search snippet.\n\nDeepMind's page, gzip-decoded in this pass, describes native audio as sound effects, ambience, and dialogue, plus ingredients or reference images, a style reference, character consistency, extend from the last second, first and last frame, outpainting, 1080p and 4K, and SynthID. Benchmark charts on that page say they were last updated October 2025 and describe 8-second clips. Those charts are not a 2026-10-16 spec. Prompt-language coverage beyond Vertex's \"English\" was not copied from DeepMind into this table.\n\n### Prompt notes\n\nNo Veo prompt was executed for this page. Vertex limits the prompt language to English on the GA card. A multi-shot storyboard inside one 8-second clip is a prompting habit, not a schema field. If the spot needs a second shot, the documented operation is extend, which the card lists separately, or a second request. Reference image to video is locked to 8 seconds. A 4-second reference-image job is outside the card.\n\nArena numbers were not opened. Not observed.\n\n## Where to run\n\n1. Vertex / Gemini Enterprise Agent Platform model card for Veo 3.1, including `veo-3.1-generate-001`: [the Veo 3.1 generate card](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/veo/3-1-generate). Use it for GA status, region, length, and capabilities. It is not a public per-second menu.\n2. Gemini API pricing: the Veo 3.1 block on [ai.google.dev/gemini-api/docs/pricing](https://ai.google.dev/gemini-api/docs/pricing), last updated 2026-10-15 UTC. This is the with-audio price list and the free-tier denial.\n3. fal: [fal-ai/veo3.1](https://fal.ai/models/fal-ai/veo3.1). This is the with-audio and without-audio price, plus the same 4/6/8 duration enum.\n4. Runway's developer price guide: [Runway API pricing](https://docs.dev.runwayml.com/guides/pricing), rows `veo3.1` and `veo3.1_fast`. The Runway product page is [/tools/runway](/tools/runway).\n5. DeepMind's public model page: [deepmind.google/models/veo](https://deepmind.google/models/veo), which points at Gemini and Flow. Flow is a Google product surface, not a model id.\n6. Google Flow, the app for Veo 3.1: [/tools/google-flow](/tools/google-flow). Plan stickers live on that tool page, verified 2026-10-17. A Flow credit is not a Gemini per-second rate.\n\nLuma's pricing page also lists Veo 3.1 in Luma credits. That is another host, noted under pricing. Higgsfield's saved catalog snapshot did not put a Veo model in the source list that was indexed for this pass. Absence from that snapshot is not proof Veo is missing from every Higgsfield account. The catalog text itself says availability can change.\n\n## Multi-channel pricing\n\n**Gemini API, paid, with audio, default.** Free tier: not available. Standard: $0.40 per second at 720p and 1080p, $0.60 per second at 4k. Fast: $0.10 at 720p, $0.12 at 1080p, $0.30 at 4k. Lite: $0.05 at 720p, $0.08 at 1080p, and 4k is not supported. The page says you are charged only if the video generates successfully. An audio-off row was not in that Veo 3.1 block. A note on the page says an audio processing issue can prevent a video in some cases. The rest of that note was not needed to price the row, and it should not be paraphrased into a success-rate claim. Verified against the HTML fetched 2026-10-16. The footer date is 2026-10-15 UTC.\n\nAn 8-second standard 1080p clip at $0.40 per second is $3.20 if you multiply. That product is derived. Lite at $0.05 for 8 seconds at 720p is $0.40 derived. Do not present the derived clip total as a second Google SKU.\n\n**Vertex GA card.** Pay-as-you-go: not supported. No dollar prices on the card. Provisioned throughput and fixed quota: supported. The \"50 tokens per minute\" quota is quoted and not converted.\n\n**fal standard `fal-ai/veo3.1`.** $0.20 per second without audio or $0.40 per second with audio at 720p or 1080p. At 4k, $0.40 without audio or $0.60 with audio. fal's example: 5 seconds at 1080p with audio on costs $2.00. The with-audio 720p/1080p figure matches Gemini standard. fal adds a silent discount Gemini's table does not list. A fal fast `llms.txt` was not saved, so fal fast prices are not observed.\n\n**Runway credits.** $0.01 per credit. `veo3.1` audio 40 credits per second, no audio 20. `veo3.1_fast` audio 15, no audio 10. Derived: $0.40, $0.20, $0.15, and $0.10 per second. The $0.40 and $0.20 derived figures match fal's standard with-audio and silent rates. The fast derived $0.15 / $0.10 do not match Gemini fast ($0.10 / $0.12 / $0.30). Host fast tiers are not Google's fast tier. Label every Runway figure derived from the credit table.\n\n**Luma credits, not dollars.** On [lumalabs.ai/pricing](https://lumalabs.ai/pricing), fetched 2026-10-16, Veo 3.1 text-to-video and image-to-video are 140 credits per second at 720p and 1080p when the audio column is not Yes, and 280 credits per second when the audio column is Yes. A USD value per Luma credit was not printed on that Veo block.\n\n## Workflow and platform fit\n\nThese are site workflow labels, not published guide URLs.\n\n**Faceless.** Eight seconds with native audio and a 9:16 option is a hook, not a full faceless episode. Plan on several generates plus an editor. English-only prompts on the Vertex card constrain non-English faceless scripts at the prompt layer. Captions still belong in [/tools/capcut](/tools/capcut). A separate narrator belongs in [/tools/elevenlabs](/tools/elevenlabs) if you do not want Veo's generated dialogue.\n\n**UGC ads.** First and last frame plus a reference image are the storyboard controls Vertex lists. Reference image to video is 8 seconds only. That is a single beat of a UGC ad, often the product hero, not the whole testimonial. A presenter with a locked face is [/models/heygen-video](/models/heygen-video) or [/tools/creatify](/tools/creatify).\n\n**AI avatar.** Veo is not an avatar product. Character consistency is a DeepMind marketing claim and a reference-image capability. It is not a photo-avatar rate card. Do not scope a spokesperson retainer onto Veo without a test you run yourself. This directory did not run that test.\n\n**Long-to-short.** Extend is the closest native operation, and it still produces another short Veo clip rather than a chaptered cutdown of a long recording. Cut the long source elsewhere, then optionally cover a beat with Veo.\n\n**Captions and cover.** C2PA and SynthID are provenance signals, not captions. A cover is a still. Veo outputs video. Make the cover in an image tool or by grabbing a frame in CapCut.\n\n**Voice and publish.** Sound generation can include dialogue on DeepMind's description. Vertex does not publish a speaker-id or a language switch. Prompt language is English. If the publish channel is non-English, treat generated speech as unverified and lay a voice track. 9:16 is supported on Vertex and on fal, which is the Shorts, Reels, and TikTok frame. 1:1 was not in the aspect list. 4K output is on the Vertex card and on the Gemini standard and fast rows. Lite cannot do 4K.\n\n**Shorts vertical fit.** Good for a vertical 4, 6, or 8 second beat with audio, at a known high per-second price on the standard tier. Weak as the only generator for a 30 to 60 second Shorts story, because you will chain extends or cuts and pay by the second each time. A 720p Lite or Fast row is the budget experiment. Standard 4K is a master, not a phone draft.\n\n## Pros\n\n- The GA card names length, aspect, frame rate, container, region, and launch stage in one place, which is rarer than a marketing grid.\n- 9:16 is supported on both Vertex and fal, so a vertical deliverable is an enum value.\n- Sound generation is a capability flag, and Gemini prices the with-audio second directly. You are not guessing whether audio is included.\n- fal publishes a silent price that the Gemini table does not, which gives a real second channel for B-roll.\n- C2PA on Vertex and SynthID on the DeepMind page are the provenance hooks a brand review can ask for. This page does not evaluate how those marks survive a re-encode in CapCut.\n- Successful-generation billing is stated on the Gemini price page, which matters when a filter rejects a prompt.\n\n## Cons\n\n- The GA id and the Gemini price ids are different strings. Quoting $0.40 per second \"for Veo 3.1 GA\" overclaims what the price page labeled. The price page labeled preview ids.\n- Eight seconds is a short ceiling for one request. Reference-to-video cannot use the 4 or 6 second options.\n- Vertex pay-as-you-go is not supported on the card that proves GA. A team that can call Gemini preview prices may still lack the Vertex SKU, and the reverse can also be true.\n- Prompt language on the GA card is English only. That is a hard line for many Shorts markets.\n- Multi-shot inside one request was not a field. Extend is a second operation with its own cost.\n- The DeepMind page still contains Veo 3 introductory sections under a 3.1 title. Cite the Vertex card or the price table when a sentence disagrees.\n- Region evidence in this pass is `us-central1` on the GA card. Other regions were not listed there. Do not invent a global allow-list.\n\n## Alternatives\n\n- [Kling](/models/kling). Kling's fal duration enum runs from 3 to 15 seconds and its official 1080p with-audio rates print below Gemini's $0.40 standard second. Kling also has a multi-shot prompt field Veo' s fal schema lacks.\n- [Seedance](/models/seedance). Seedance 2.5 reaches 30 seconds on the fal US card and on the Higgsfield image-to-video schema, with audio included in the fal video price. Veo is the shorter, Google-contracted option with a GA region line.\n- [Runway Gen](/models/runway-gen). Gen-4.5 is Runway's own 12-credit second. Veo on Runway is the `veo3.1` credit row. The names share a wallet and not a model.\n- [Luma Ray](/models/luma-ray). Ray3.2 text-to-video has no native audio, longer modify paths, and HDR/EXR. Veo is the better default when the clip must contain generated sound and a C2PA mark is on the spec card.\n- [MiniMax Hailuo](/models/minimax-hailuo). Hailuo is a general T2V/I2V alternative with Luma and other host wallets. Veo is the better default when you need Google-contracted audio and a C2PA mark on the GA card.\n\n## Related tools\n\n- [/tools/google-flow](/tools/google-flow) is Google's app for Veo 3.1. Plan stickers are on that page, verified 2026-10-17. This model page keeps the Gemini and Vertex rates.\n- [/tools/runway](/tools/runway) is the credit host that publishes Veo rows beside Gen-4.5.\n- [/tools/capcut](/tools/capcut) is the editor for captions, covers, and stitching 8-second beats into a Short.\n- [/tools/elevenlabs](/tools/elevenlabs) is the controlled voice when English-only generated dialogue is the wrong asset.\n- [/tools/higgsfield](/tools/higgsfield) is a multi-model studio. A Veo endpoint was not in the catalog index saved for this pass. Confirm inside the console before you promise that route.\n- [/tools/creatify](/tools/creatify) and [/tools/heygen](/tools/heygen) cover ad templates and presenters when Veo's 8-second world clip is the wrong job.\n- [/tools/kling](/tools/kling) and [/tools/seedance](/tools/seedance) are the product apps for the closest alternative models.\n\n## FAQ\n\n### Is Veo 3.1 generally available?\n\n`veo-3.1-generate-001` is GA on the Vertex card, in `us-central1`, with a retirement footnote of November 17, 2026 or later for that platform. The Gemini API price page lists preview model ids and does not offer a free tier. Say which id you mean.\n\n### Does the $0.40 rate include audio?\n\nOn the Gemini API page, $0.40 is the standard with-audio price at 720p and 1080p. fal charges $0.40 with audio and $0.20 without at those resolutions. Vertex's card does not print the dollar rate.\n\n### Can I make a 9:16 Short?\n\nYes, 9:16 is on the Vertex aspect list and on the fal schema. Length is still 4, 6, or 8 seconds per request, and a reference-image request is 8 seconds only.\n\n### Is extend the same as multi-shot?\n\nNo. Extend is a documented separate capability. The fal schema used here has no multi-shot prompt array. The filter on this page sets multi-shot to false for that reason.\n\n### Where are the benchmark scores?\n\nDeepMind charts on the decoded page are labeled last updated October 2025. They are not reused as a 2026-10-16 quality rank. Artificial Analysis was not opened. Current arena scores are not observed.", "path": "/models/veo" } ] }