Every “best AI video generator” list you have opened this week was written for somebody buying a monthly subscription. This one is for somebody who has to call one. Twelve models are worth a place on your shortlist in September 2026, and not one of them is best at everything: Vidu Q3 Pro, Wan 3.0, PixVerse V6, Hailuo 2.3, Seedance 2.0 Fast, Kling V3, MiniMax H3, Grok Imagine Video, SkyReels, Seedance 2.5, HappyHorse 1.1 and Veo 3.1. They come from nine providers, and their listed prices run from $0.0450 to $0.5400 per second of generated video — a spread of roughly 12× for the same nominal job.
Modellix runs this blog and distributes all twelve routes, so we have a commercial interest in this page. That is why every price below is a quote you can check rather than a claim you have to trust: all twelve were read from the live model pages on September 15, 2026, one URL per row, plus the exact request body each route expects. Nothing here is ranked by how good the output looks. The order is price, and we say so where the table starts.
If what you actually wanted was a business-facing roundup of subscription tools — Canva, HeyGen, Synthesia and the rest — our AI video generators roundup for business teams is the page for that question. This page answers a different one: which model do I put in my code, what does it cost per call, and what does the request look like.
The twelve best AI video generators, at a glance
Ordered by listed floor price, cheapest first. “Max clip” and “Max resolution” are the ceilings published in each model’s own request schema, not our summary of them.
| Model | Provider | List price / second | Max clip | Max resolution |
|---|---|---|---|---|
vidu/viduq3-pro-t2v |
Vidu | $0.0450~$0.1200 | 16 s | 1080p |
alibaba/wan3.0-t2v |
Alibaba | $0.0500~$0.2000 | 30 s | 1080P |
pixverse/v6-t2v |
PixVerse | $0.0500~$0.2300 | 15 s | 1080p |
minimax/hailuo-2.3-t2v |
MiniMax | $0.0504~$0.0738 | 10 s | 1080P |
bytedance/seedance-2.0-fast-t2v |
ByteDance | $0.0600~$0.1200 | 15 s | 720p |
kling/kling-v3-t2v |
Kling | $0.0672~$0.3360 | 15 s | 4k |
minimax/minimax-h3-t2v |
MiniMax | $0.0800~$0.1300 | 15 s | 2K |
xai/grok-imagine-video |
xAI | $0.0840~$0.1080 | 15 s | 720p |
skywork/skyreels-t2v |
Skywork | $0.0960~$0.4200 | 15 s | 1080p |
bytedance/seedance-2.5-t2v |
ByteDance | $0.1030~$0.5690 | 30 s | 1080p |
alibaba/happyhorse-1.1-t2v |
Alibaba | $0.1400~$0.1800 | 15 s | 1080P |
google/veo-3.1-t2v |
$0.3600~$0.5400 | 8 s | 4k |
Listed per-second price range for each of the twelve routes, read from the model pages on September 15, 2026. The dark segment is the floor of the published range; the amber segment is how far the ceiling sits above it.
The table is one column short of the full picture, and the missing column is the one that decides most of these comparisons: audio. Six routes expose a switch — Kling V3 takes audio: native|off, Wan 3.0 audio, Seedance 2.0 Fast and Seedance 2.5 generate_audio, PixVerse V6 generate_audio_switch, SkyReels sound. Three more ship audio natively with no control parameter at all: Veo 3.1 documents synchronised audio as part of the video, and Vidu Q3 Pro and HappyHorse 1.1 both advertise native audio in their model descriptions while exposing nothing in the schema. The remaining three — Hailuo 2.3, MiniMax H3 and Grok Imagine Video — publish no audio field anywhere. If your product needs sound inside the file, that narrows this list from twelve to nine before price gets a say.
The most expensive route on the shortlist, on its own model page on September 15, 2026. The badge is where every number in the table above came from — it is on the page, not in this article’s data.
Floor prices are not the prices most of these routes actually charge you, and each model page publishes its own tier table. Read on the same day, at 1080p — a resolution most of this list supports — the tiers run from $0.0738/second on Hailuo 2.3 (which is the lowest 1080p tier of the twelve) through $0.0896 on Kling V3 with audio off, $0.1200 on Vidu Q3 Pro, $0.1800 on HappyHorse 1.1 and PixVerse V6, $0.2000 on Wan 3.0, $0.3300 on SkyReels in fast mode and $0.5690 on Seedance 2.5. Veo 3.1 charges the same $0.3600 at 720p and 1080p and only steps up to $0.5400 at 4k. Three routes publish no 1080p tier and cannot be placed on this axis: Seedance 2.0 Fast and Grok Imagine Video cap at 720p, and MiniMax H3’s tiers are 768P and 2K at $0.1300 — it jumps straight over 1080p. That is a 7.7× spread at the resolution you were probably going to use — which is the whole reason this list is ordered by floor and priced by tier.
How this shortlist was built
Modellix’s video catalogue held 112 video routes across nine providers when we read the model pages on September 15, 2026, priced from $0.0050 to $1.4400 per second. Twelve from 112 needs a rule, so here is the one we used.
One flagship per provider, plus the families where a second tier is genuinely a different purchase. Nine providers means at least nine entries; the other three are ByteDance Seedance 2.0 Fast (a cheaper, 720p-capped sibling of Seedance 2.5), MiniMax Hailuo 2.3 (a differently priced sibling of MiniMax H3) and Alibaba HappyHorse 1.1. That is how 12 comes out of nine, rather than twelve brands you have heard of.
Task routes were excluded. The 112 includes lip-sync, video upscaling, video editing, restyling and digital-human routes. They produce video, but they do not answer “best AI video generator” — nobody shortlists a lip-sync model for a text prompt. They are in the catalogue and on the model index if that is the job you have.
Nothing here is a quality ranking. We have not run a side-by-side prompt test for this page, and we are not going to imply one we did not do. What the twelve entries below give you is the price, the parameters and one honest constraint per model — the three things you need before you spend an afternoon testing. If you want a second opinion on which models belong on a shortlist at all, fal.ai’s roundup of AI video generators is the closest thing the first page has to an API-layer view, and the model-selection guide in our own docs groups the catalogue by the job rather than by the provider.
The catalogue the shortlist was drawn from, captured September 15, 2026. The index lists every route by provider and model id; the twelve above are a selection from it, not the whole page.
How to call any of them: submit, poll, retrieve
This is the part that is the same for all twelve, and the part the rest of the first page for this keyword does not cover — a scan of the four ranking pages we read for this article turned up no endpoint, no API key and no async task handling anywhere in 8,513 words of text.
Authentication is one header, and the key comes from the console. The full lifecycle is documented in Modellix’s REST API guide, and it starts like this:
1 | curl --request POST \ |
Every model call on that host is an asynchronous task, so the response is not your video — it is a handle for it:
1 | { |
Keep the task_id and poll the URL the response hands you; the task-result endpoint reference covers the full response. When status flips to success, the assets arrive under result.resources, each with its own url, type, width, height, format and role. A minimal Python loop — the shape is the same for all twelve models, only the body changes:
1 | import os, time, requests |
Three details worth knowing before you build on this. Generated assets are retained for seven days, so copy them to your own storage as part of the pipeline rather than after it. Polling is optional: send an X-Webhook-URL header on the submit call and the platform posts the same payload to your HTTPS endpoint when the task reaches a terminal state, retrying on 429 and 5xx and not retrying on 4xx — which is the version you want in production. And image inputs go through a separate File API (POST /api/v1/media/files, multipart/form-data with the field named file), which returns a URL you pass into the model’s image field; uploads are not billed, but they are capped at 16 MB per file and about seven days of retention.
Errors are worth reading once rather than logging forever: 400 means fix the parameters, 401 a bad key, 402 an empty balance, 404 a task or model that does not exist, 429 you are over concurrency or RPM — check X-RateLimit-Reset and back off — and 500/503 are the two codes worth retrying with exponential backoff. Concurrency and RPM scale with your funding tier rather than being a fixed number, and the entitlements page has the current table.
Video model API reference
Every route's request schema, parameter enums and poll endpoint in one place.
View Docs
The submit half of the lifecycle, captured September 15, 2026: the POST to a model’s /async endpoint and the task_id it returns. The get_result.url in that response is the only polling address you need.
The retrieve half: a completed task returns its assets under result.resources, and the response carries the billed amount. This is the same payload a webhook delivers.
The twelve models, one by one
Each entry gives you the route, the price read on September 15, 2026, the parameters its schema actually accepts, and one constraint that is worth knowing before you commit. The request bodies are transcribed verbatim from each model’s published OpenAPI schema — the same source as the parameter lists — and are meant to be run against your own key; they have not been executed against a live account for this page, so treat the first one as a smoke test of your own credentials rather than a verified call.
1. Vidu Q3 Pro T2V — vidu/viduq3-pro-t2v
Vidu’s premium cinematic text-to-video route, and the cheapest floor on this list at $0.0450~$0.1200 per second. The schema is deliberately small: prompt-only, with duration, aspect ratio, resolution and an optional seed for reproducibility.
1 | curl --request POST --url https://api.modellix.ai/api/v1/vidu/viduq3-pro-t2v/async \ |
Required: prompt. Optional: duration, aspect_ratio (16:9, 9:16, 1:1, 3:4, 4:3), resolution (540p, 720p, 1080p), seed. Constraint: the badge on this route publishes a range, not a unit price, and resolution is one of the parameters that moves you along it — treat $0.0450/second as the price of your cheapest acceptable settings, not of your intended ones. Vidu’s own model page also states that this route does not support automatic background music and points you at Q3 Turbo T2V if you need fast previews. Model page
The cheapest floor on the shortlist, read from its model page on September 15, 2026. The badge publishes a range rather than a unit price; resolution (540p, 720p or 1080p) is one of the parameters that decides where on it you land.
2. Wan 3.0 T2V — alibaba/wan3.0-t2v
Alibaba’s Wan 3.0 is the longest-clip option here alongside Seedance 2.5: 2 to 30 seconds in a single call, at $0.0500~$0.2000 per second. It is also unusual in accepting three different starting points — a text prompt, an uploaded file, or a public webpage link — for grounding the generation.
1 | curl --request POST --url https://api.modellix.ai/api/v1/alibaba/wan3.0-t2v/async \ |
Required: one of prompt, file_url or link_url — this route will not accept an empty body. Optional: resolution (480P, 720P, 1080P), ratio (16:9, 4:3, 1:1, 3:4, 9:16), duration (2–30), audio (defaults to true), seed. Constraint: audio is on by default, so if you want a silent file you have to turn it off explicitly. Alibaba documents the same route family in its Wan text-to-video API reference if you want the provider’s own view of the parameters. Model page
3. PixVerse V6 T2V — pixverse/v6-t2v
PixVerse V6 publishes eight aspect ratios — including the 2:3/3:2 pair that only Grok Imagine Video also carries here — and is one of two routes with a multi-clip toggle (generate_multi_clip_switch; Kling V3’s multi_shot is the other). It is priced at $0.0500~$0.2300 per second.
1 | curl --request POST --url https://api.modellix.ai/api/v1/pixverse/v6-t2v/async \ |
Required: prompt, aspect_ratio, duration, quality — four required fields, matched on this list only by MiniMax H3. aspect_ratio accepts 16:9, 9:16, 1:1, 3:4, 4:3, 2:3, 3:2, 21:9; quality maps to 360p, 540p, 720p or 1080p; duration is an integer from 1 to 15. Optional: seed, generate_audio_switch, generate_multi_clip_switch. Constraint: because duration and quality are mandatory, there is no “just send a prompt” path here — a first smoke test needs four parameters, not one. Model page
4. Hailuo 2.3 T2V — minimax/hailuo-2.3-t2v
Hailuo 2.3’s published range is a narrow one — $0.0504~$0.0738 per second, a 1.46× spread — and its 1080P tier at $0.0738 is the lowest 1080p price of the twelve. MiniMax positions this model for human performance and stylised aesthetics rather than general scenery.
1 | curl --request POST --url https://api.modellix.ai/api/v1/minimax/hailuo-2.3-t2v/async \ |
Required: prompt. Optional: prompt_optimizer (defaults to true), fast_pretreatment, duration, resolution. Constraint: duration is not a range — it accepts 6 or 10 seconds and nothing in between — and the two parameters interact: the schema states it twice, that a 10-second clip only supports 768P. A 1080P Hailuo clip is a 6-second clip. Model page
5. Seedance 2.0 Fast T2V — bytedance/seedance-2.0-fast-t2v
ByteDance’s faster Seedance variant, at $0.0600~$0.1200 per second — the cheapest way onto the Seedance family. It also carries more runtime knobs than most: a camera lock, an audio toggle, a last-frame return and an execution expiry.
1 | curl --request POST --url https://api.modellix.ai/api/v1/bytedance/seedance-2.0-fast-t2v/async \ |
Required: prompt. Optional: ratio (16:9, 4:3, 1:1, 3:4, 9:16, 21:9, adaptive), resolution (480p or 720p), duration (4–15), seed, camera_fixed, generate_audio, execution_expires_after, return_last_frame. Constraint: the fast variant caps at 720p — if 1080p is a requirement, this is the wrong sibling and Seedance 2.5 is the one you want. ByteDance’s own ModelArk documentation is the provider-side entry point for this family. Model page
6. Kling V3 T2V — kling/kling-v3-t2v
Kling V3 is one of only two routes here that reach 4k, priced at $0.0672~$0.3360 per second — a 5× spread between its 720p/no-audio tier and its 4k tier. It documents native synchronised audio and multi-shot generation directly from text.
1 | curl --request POST --url https://api.modellix.ai/api/v1/kling/kling-v3-t2v/async \ |
Required: prompt. Optional: audio (native or off), multi_shot, resolution (720p, 1080p, 4k; defaults to 720p), aspect_ratio (16:9, 9:16, 1:1; defaults to 16:9), duration (3–15; defaults to 5). Constraint: the aspect-ratio enum is only three values wide — no 4:5 or 21:9 crop here. Kling publishes its own developer API documentation if you need the provider-side reference. Model page
Kling V3’s badge, captured September 15, 2026. Its published range spans 5× between floor and ceiling, and the two parameters that move you along it are resolution (720p, 1080p, 4k) and audio (off or native).
7. MiniMax H3 T2V — minimax/minimax-h3-t2v
MiniMax H3 is the second MiniMax entry and the more demanding of the two by design, at $0.0800~$0.1300 per second with a 1.62× spread.
1 | curl --request POST --url https://api.modellix.ai/api/v1/minimax/minimax-h3-t2v/async \ |
Required: prompt, duration, resolution, ratio — all four. duration runs 4–15, resolution is 768P or 2K, ratio accepts 21:9, 16:9, 4:3, 1:1, 3:4, 9:16. Constraint: the schema marks all four as required even though it also lists a default for three of them, so send all four rather than relying on a fallback. MiniMax’s pay-as-you-go pricing page is the provider-side reference for this family. Model page
8. Grok Imagine Video — xai/grok-imagine-video
xAI’s text-to-video route is one of the two most predictable prices on the list — $0.0840~$0.1080 per second, a 1.29× spread, tied with HappyHorse 1.1 — and the simplest schema: prompt, duration, aspect ratio, resolution.
1 | curl --request POST --url https://api.modellix.ai/api/v1/xai/grok-imagine-video/async \ |
Required: prompt. Optional: duration (1–15, defaults to 8), aspect_ratio (1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3), resolution (480p or 720p). Constraint: 720p is the ceiling. Price stability is real, but it is bought with a resolution cap that the price itself does not tell you. Model page
9. SkyReels T2V — skywork/skyreels-t2v
Skywork’s SkyReels route spans $0.0960~$0.4200 per second and is unusual in exposing an explicit quality-versus-speed switch rather than hiding it behind a resolution enum.
1 | curl --request POST --url https://api.modellix.ai/api/v1/skywork/skyreels-t2v/async \ |
Required: prompt. Optional: duration (3–15, defaults to 5), aspect_ratio (16:9, 4:3, 1:1, 9:16, 3:4), resolution (480p, 720p, 1080p), mode (fast or std), sound, prompt_optimizer. Constraint: mode is a separate axis from resolution, so a 1080p fast run and a 1080p std run are different jobs at the same nominal resolution — pin both if you are comparing cost between runs. Model page
10. Seedance 2.5 T2V — bytedance/seedance-2.5-t2v
Seedance 2.5 is the current ByteDance flagship: 4 to 30 seconds, up to 1080p, native audio, and a choice of mp4 or mov output, at $0.1030~$0.5690 per second — a 5.52× spread, the widest on this list.
1 | curl --request POST --url https://api.modellix.ai/api/v1/bytedance/seedance-2.5-t2v/async \ |
Required: prompt. Optional: ratio (16:9, 4:3, 1:1, 3:4, 9:16, 21:9, adaptive), resolution (480p, 720p, 1080p), duration (4–30), seed, camera_fixed (default false), generate_audio (default true), output_format (mp4 or mov), execution_expires_after, return_last_frame (default false). Constraint: the 5.52× spread is the whole story of this model’s pricing — a 30-second 1080p run is not the same purchase as a 4-second 480p one, and both are “Seedance 2.5”. Model page
Seedance 2.5’s badge, captured September 15, 2026. This route publishes the widest range on the shortlist — 5.52× between its floor and its ceiling.
11. HappyHorse 1.1 T2V — alibaba/happyhorse-1.1-t2v
Alibaba’s HappyHorse 1.1 is the newest route on the list and one of the two tidiest to forecast, a 1.29× spread: $0.1400~$0.1800 per second, 720P or 1080P, 3 to 15 seconds.
1 | curl --request POST --url https://api.modellix.ai/api/v1/alibaba/happyhorse-1.1-t2v/async \ |
Required: prompt. Optional: resolution (720P or 1080P), ratio (16:9, 9:16, 1:1, 4:3, 3:4, 4:5, 5:4, 9:21, 21:9), duration (3–15), seed. Constraint: HappyHorse 1.0 remains on the catalogue at a different price, so pin the version in code rather than relying on a family name. Model page
12. Veo 3.1 T2V — google/veo-3.1-t2v
Google’s Veo 3.1 is the most expensive route on the shortlist at $0.3600~$0.5400 per second, and the only one of the twelve whose schema documents both a personGeneration policy and a separate negativePrompt field. It reaches 4k with natively synchronised audio.
1 | curl --request POST --url https://api.modellix.ai/api/v1/google/veo-3.1-t2v/async \ |
Required: prompt. Optional: negativePrompt, aspectRatio (16:9 or 9:16; defaults to 16:9), duration (4, 6 or 8), resolution (720p, 1080p, 4k), personGeneration. Constraint: the schema says it outright — 1080p and 4k are only available at an 8-second duration. A 4-second Veo clip is a 720p clip billed at the 1080p rate: $0.3600 per second for 720p output. Google’s own video generation guide documents the underlying model family. Model page
Every one of the twelve also exists as an image-to-video or reference-to-video route on the same provider — the mode changes, the per-second economics do not stay identical, and we have deliberately printed only the text-to-video prices we read and could check on September 15, 2026. Our API-level comparison of models published earlier this year takes the same routes from a different angle, and the full text-to-video model list is where to go if twelve is too narrow a shortlist.
What a per-second price does and does not tell you
It does not tell you the top of the range. Of the 112 video routes in the catalogue, 88 publish a range rather than a single number, and the twelve here are no exception: only the price floor is a fixed, checkable quantity. The ceiling moves with the parameters you send — resolution, duration, mode — which is why Kling V3’s ceiling is 5× its floor and Seedance 2.5’s is 5.52×. A budget built on floor prices is a budget for minimum quality settings.
It does not tell you what a clip costs. Per-second is a unit rate, and unit rates multiply. Worked at the floor price of each route, and rounded to the cent:
Cost of one minute of generated video at each route’s listed floor price, September 15, 2026. Arithmetic on published numbers: floor price × 60. It is a floor, not a forecast — the ceiling is up to 5.5× higher on some of these routes.
One minute of output costs $2.70 on Vidu Q3 Pro’s floor and $21.60 on Veo 3.1’s — the same eight-fold gap you see per second, unchanged by scale. That is the number to put in front of whoever owns the budget, because “per second” means nothing to them and “eight times” does.
It does not tell you what the cheapest thing in the catalogue is. The $0.0050 floor belongs to digital-human routes, not to a general-purpose generator, and the $1.4400 ceiling belongs to a video-editing route. Both are real, both are in the model index, and neither is a text-to-video price. Quoting either as “the price of AI video” is the most common way this data gets misused.
What it does tell you, reliably: the price of the call you are about to make. Every route here is billed per second of output rather than per month or per credit, which means the cost of a test is bounded and small — Hailuo 2.3’s shortest clip is six seconds for about $0.30, and Veo 3.1’s is four seconds for about $1.44. That is the actual argument for pricing a shortlist by trial rather than by review: you can afford to run the same prompt across all twelve and look at the files.
One key across the nine providers behind these routes: what it changes, and what it does not
Nine providers means nine sets of credentials, nine balances, nine error formats and nine async lifecycles if you integrate them separately. Aggregating them means one API key, one bill and one submit-poll-retrieve contract, and switching a model becomes a URL and a request body rather than a new vendor integration — which is the same argument we made in running multiple video models on one API. That is a workflow claim about the shape of your integration, and it is the only claim we are making.
To be explicit about what that is not: this is not a claim that Modellix is the cheapest way to reach any of these twelve models. Several of them are sold directly by the providers that make them, at prices this page does not attempt to undercut or compare. This is also not a claim about which model looks best — we did not run that test, and a shortlist ranked by anything other than price would need one. And it changes nothing about the models themselves: the resolution caps, duration ceilings and parameter requirements documented in the twelve entries above are the provider’s, not ours, and they travel with the model whichever route you call it through.
The one place the aggregation argument becomes concrete rather than abstract is the log. A per-second price is only useful if you can see what you actually spent, and the request-logs endpoint returns the billed amount per task alongside the model, the timestamp and the duration — so a cost-per-usable-clip number is arithmetic on your own data rather than an estimate from a pricing page.
How to choose
There is no winner, so the honest answer is a routing table. Pick the constraint that matters for your product and start there; if none of these rows is quite your job, browse the whole video-generator catalogue rather than forcing a fit.
| If your main constraint is | Start with | Because |
|---|---|---|
| Lowest cost per second at 1080p | minimax/hailuo-2.3-t2v |
$0.0738/sec at 1080P — the lowest 1080p tier of the twelve |
| Lowest cost per second at any resolution | vidu/viduq3-pro-t2v |
$0.0450 is the 540p tier — the lowest published floor of the twelve |
| Clips longer than 15 seconds | bytedance/seedance-2.5-t2v or alibaba/wan3.0-t2v |
Both publish a 30-second ceiling |
| 4k output | google/veo-3.1-t2v or kling/kling-v3-t2v |
The only two with 4k in the resolution enum |
| Audio inside the file | kling/kling-v3-t2v or google/veo-3.1-t2v |
Kling’s audio switch and Veo’s natively synchronised audio; Wan 3.0 and Seedance 2.5 expose generate_audio too (see above) |
| Predictable cost per call | alibaba/happyhorse-1.1-t2v or xai/grok-imagine-video |
The two narrowest ranges on the list at 1.29×; Hailuo 2.3 is next at 1.46× |
| Unusual aspect ratios | alibaba/happyhorse-1.1-t2v or pixverse/v6-t2v |
Nine and eight values; PixVerse is one of the two routes carrying 2:3/3:2 |
| A first test today | xai/grok-imagine-video |
Prompt, duration, ratio, resolution — four fields, one required |
The per-model pages linked in the twelve entries above are the authoritative source for every price on this page. If you want the single-provider breakdown instead, our Veo 3.1 price breakdown and Kling 3.0 cost breakdown both exist and go deeper on one family each.
Tools, templates and licences: what this page does not price
Twelve routes is a shortlist of models, and most people searching this phrase are not shopping for a model. So here is the honest map of what sits either side of that line, without pretending we price any of it.
| If what you need is | Where that job lives | What this page can tell you |
|---|---|---|
| A finished video with an editor around it | Canva, InVideo AI and similar end-to-end tools | These are models behind a UI; a subscription price is not a per-second price, and the two cannot be divided into each other |
| A talking presenter or a localised avatar | HeyGen, Synthesia | Digital-human and avatar routes do exist in this catalogue — Vidu, Skywork and Kling all sell them — but they are task routes, not the twelve text-to-video models above |
| A vertical 9:16 cut for TikTok, Reels or Shorts | Most of the twelve | This is the aspect_ratio / ratio parameter: eleven of the twelve accept 9:16, and the enums are listed in each entry |
| Commercial rights you can show a client | The provider’s own terms of service | An API unit price says nothing about licensing, and nothing on this page settles it |
| A voiceover or a narration track | — | None of the twelve routes takes a voiceover parameter. A voice track is a separate model call, and this page does not price one |
Four names you may have expected are deliberately absent, and not by oversight: Runway, Pika, Luma Dream Machine and OpenAI Sora are not among the routes in this catalogue, so this page does not quote a price for any of them, and we are not going to fill the gap with a third-party estimate. If one of those is what you already use, that is a legitimate choice; it is simply a different comparison.
Frequently asked questions
Is there a 100% free AI video generator?
Not on this list, and not for callable API access. None of the twelve routes has a free tier; the catalogue is pay-as-you-go with no monthly minimum, and new registrations no longer receive a signup credit — trial credit is requested by email. The genuinely free path to AI video is running an open-weight model on your own hardware, which trades the API bill for a GPU bill and an operations bill. If your requirement is literally zero dollars, self-hosting is the answer and this page is not.
Can ChatGPT generate videos for free?
That question is about a consumer product, and nothing here prices it. What this data can tell you is narrower and checkable: OpenAI is not one of the nine providers behind the 112 video routes in this catalogue, so if you want a callable video model, the names above are the ones with published per-second prices to compare.
How much does one clip actually cost?
Multiply the listed per-second price by the duration you request. Worked at the floor prices read on September 15, 2026, and at a duration each route actually accepts: Kling V3 at 5 seconds is about $0.34 and Seedance 2.5 at 5 seconds about $0.52, while Hailuo 2.3 allows only 6 or 10 seconds — its shortest clip is about $0.30 — and Veo 3.1 allows only 4, 6 or 8, so its shortest is about $1.44 and its 8-second 1080p clip about $2.88. At each route’s ceiling it costs more — up to 5.5× more on Seedance 2.5 — because the ceiling applies when you ask for higher resolution or longer output.
Which model produces the most realistic video?
Realism is not a parameter, so no schema can answer this and we are not going to imply a test we did not run. What the schemas do give you is the levers that correlate with it: resolution, duration and native audio. Our practical suggestion is the one this pricing model makes cheap — run the same prompt across three or four candidates, look at the files, and let the result decide. A comparison built on somebody else’s screenshots is a comparison built on their prompt.
Why do the prices show a range instead of one number?
Because the parameter that changes the work changes the price. 88 of the 112 video routes in the catalogue publish a range rather than a flat figure, and for the twelve above, resolution, duration and the mode switch are what move you along it. The floor is reachable — at the cheapest settings the route allows. If you are budgeting, budget the ceiling for the settings you actually intend to send.
Do I need a separate account for each model?
Not through this route. One key reaches all nine provider brands behind these video routes — the count on the model pages, not the wider catalogue’s brand total — and every model on the list is called with the same submit-poll-retrieve pattern shown earlier in this article; what changes between them is the URL and the body. The twelve models are also available individually from the providers that make them, which is a legitimate choice if you only ever need one.
Is there an app, or does this work from a phone?
No. There is no consumer app for these routes — they are REST endpoints, and the output lands on a URL that expires after about seven days unless you store it. If you want a phone app that generates video from a prompt, you want a different product than an API aggregator.
Model availability and per-second pricing reflect public model pages read on September 15, 2026, and change frequently — validate against each model’s live page before committing. Access video and image models, including the leading Chinese models, through a single API key at modellix.ai.
