AI Video Model API Pricing: September 15 Snapshot of 112 Models

Availability update, September 24, 2026: Modellix removed ten Skywork routes after this price snapshot was captured. The route IDs remain below as historical data only; use the live video model catalog to check what you can call now.

This is a dated per-second snapshot of the 112 video routes listed on September 15, 2026, including the prices displayed on their model pages and the billing differences that make those numbers hard to compare directly. At the time, the routes came from nine provider catalogs. Modellix runs this blog and distributed those routes, so we have a commercial interest in this comparison; every listed rate was visible on its linked model page on that date.

The September 15 snapshot: 112 routes ranged from $0.0050 to $1.4400 per second. The median lowest listed rate was $0.0710/sec, or about $0.36 for a five-second clip. 88 routes published a range because price changed with resolution or task type; the top of one model's range could be 11.1× its floor.

What 112 video routes cost in the September 15, 2026 snapshot

Start with the shape of the catalog. The 112 routes split by what you hand the model, and the cheapest and most expensive point in each group is a different model:

Every row above is a model page URL of the form https://www.modellix.ai/models/<model_path>. The cheapest number in the catalog belongs to a digital-human route, not a general-purpose generator, a reminder that "video model" covers lip-sync, avatar, upscaling and video editing, and those tasks price very differently from a five-second cinematic clip.

The ceiling is worth looking at for the same reason, because it makes the resolution mechanic visible before the tables get to it:

Modellix model page for alibaba/happyhorse-1.0-video-edit, captured September 15, 2026: the highest per-second list price in the 112-model catalog. The resolution selector reads 1080P in the input panel below the badge, which is what puts this route at the top of its own range rather than the bottom.

The spread by provider is just as wide, and it does not sort the way people expect. Cheapest model per provider, on the same September 15 reading:

Modellix model page for vidu/viduq2-pro-digital-human, captured September 15, 2026. The outlined badge is the exact field every per-second figure in this article was read from. Note what sits below it: a resolution selector, the parameter that turns a single price into the ranges in the tables above.

Why per-second pricing is not a unit you can subtract

The reason a video price table is harder to shop than a token table is that four different billing units are in circulation, and the SERP mixes them without warning. Each one answers a different question, and only one of them is the unit in the tables above:

fal.ai's own pricing page puts the distinction plainly: video models there are billed by output unit, per second or per video, depending on the model (fal.ai/pricing, read September 15, 2026). That single sentence is the honest description of the market, and it is also why a table that mixes the two units cannot be used as a budget.

There is a fifth unit that is not a unit at all, and it derails more budgets than any of the four: the monthly seat. Consumer video tools sell a subscription, so a reader who is comparing "the API" against "the tool I already pay for" is comparing a per-second bill against a flat monthly fee. Those only become comparable at one point (the volume where your monthly billed seconds cost exactly what the seat costs), and below that volume the subscription usually wins on paper. Which of them is actually cheaper depends on how many of your generated clips you would ship, and that is a question the seat plan's price cannot answer either.

Flat versus parameter-driven pricing across the 112 video models, measured September 15, 2026. Text-to-video is almost entirely parameter-driven (25 of 26 routes publish a range), while video-to-video is nearly split down the middle, because lip-sync and upscaling tasks do not have a resolution dial to turn.

The practical consequence: when a page puts a video model and an image model in one "cost per generation" column, the column is doing arithmetic on units that do not share a denominator. We saw exactly that pattern in one of the pages ranking for this query: per-generation pricing and per-second billing presented side by side in a single table without a conversion rule (Renderful's comparison, read September 15, 2026). It is a readable table. It is not a comparable one.

And a second, subtler trap: 88 of the 112 routes publish a range, not a price. A page that quotes one number for bytedance/seedance-2.0-t2v is quoting either its floor or its ceiling and telling you neither. The range is the honest unit for these models, because the number moves with a parameter you control.

The price spectrum inside one model: resolution is the multiplier

Here is the parameter that moves a bill the most, in the catalog we measured: resolution. It is not a small adjustment. Ranking the 112 routes by how far apart their own floor and ceiling sit:

None of the competing pages we read gave this multiplier. They say resolution affects price. True, and unusable as a budget input on its own. The usable version is the ratio: the same prompt, same model, same provider, and a 4.0× to 11.1× swing on one dropdown. If your budget assumes the floor price, you are budgeting for the lowest resolution you can ship, and nothing else.

The same logic applies downward, which is where the savings live. google/veo-3.1-lite-t2v prices at $0.0450~$0.0720/sec, a 1.6× spread, narrow enough that a forecast built on its floor stays close to the actual bill.

Per-second list price span for each provider's video catalog, measured September 15, 2026. The horizontal scale is logarithmic because the spans are: Alibaba's own catalog runs from its cheapest Wan 3.0 route to happyhorse-1.0-video-edit at $1.4400/sec, a 28.8× multiple inside one provider.

One more detail from that snapshot: 24 routes had a single flat per-second price, mostly lip-sync, upscaling and avatar tasks where resolution was not the variable. skywork/sky-lipsync at $0.0120/sec was one; it was removed from Modellix on September 24. pixverse/upscale-video at $0.0500/sec was another. Flat is not the same as cheap, and a minimum duration can make a flat rate cost more per usable clip than a ranged rate at its floor.

Stack the 112 floor prices against each other and the catalog has a clear centre of gravity, which is useful for a first sanity check on any quote you are handed:

Where the 112 video models sit by cheapest published list price, measured September 15, 2026. The largest single group (52 routes) floors between $0.05 and $0.10 per second, which is why a five-second clip at the median lands near 36 cents rather than near the catalog's headline floor.

A worked example: 5-second clips, 10,000 a month

Related searches for this query are full of calculator language: people want to end up with a number, not a table. So take a concrete volume: 5-second clips, 10,000 of them a month. That is 50,000 seconds of output, and the arithmetic is one multiplication once you know the unit:

Read that table as a budgeting exercise, not a ranking. The routes do different jobs, and the digital-human row is cheap because it animates a face, not because it is a better deal on cinematic footage. What the table does show is the size of the decision: at the same volume, the range across the catalog is $250 to $72,000 a month. Nobody picks a video model by the headline rate; they pick it by which column they can live in.

At one fixed workload (5-second clips, 10,000 of them a month), the same catalog spans $250 to $72,000. The per-second rate is not the decision. Which column you can live in is the decision.

Two levers move that column faster than switching providers:

• Drop the resolution. The multiplier from the previous section applies directly: the same route at its floor instead of its ceiling can cut a line item by more than half.
• Cap the duration. A 5-second clip is one price; the same model at 8 seconds is 60% more per clip before any resolution change. Duration and resolution multiply, they do not add.

Add the multiplier most calculators leave out: attempts

The table above assumes every generation ships. It will not. The full formula has three inputs, and only one of them is a price:

cost = tier rate ($/sec) × seconds you keep × attempts per kept second

A 10-second shot that takes four generations bills 40 seconds, not 10. Read the rate from the tier you will actually render, which each model page lists under its Pricing heading. For bytedance/seedance-2.0-t2v on September 15, 2026, that was 480p $0.0700, 720p $0.1500, 1080p $0.3700 and 4k $0.7800 per second. The tier axis is not always resolution (kling/kling-avatar tiers on mode: std $0.0448, pro $0.0896), and more resolution is not always more money: google/veo-3.1-t2v charges $0.3600 at both 720p and 1080p.

Now apply it to a 60-second finished video: 12 shots × 5 seconds × 3 attempts = 180 billed seconds.

A 2-minute video on the same shape is 360 billed seconds: $25.20 on Seedance 2.0 at 480p, $129.60 on Veo 3.1 at 720p. The attempt count is the input you control: on Seedance 2.0 at 720p, the same 60-second cut costs $9.00 at one attempt per shot and $54.00 at six. Published estimates assume anywhere from three to ten attempts per usable shot, so treat the number as your own assumption and measure it.

One accounting detail worth knowing before you compare our numbers to an invoice: Modellix's 10% first-top-up discount is a checkout discount on your first purchase of each top-up tier, an account-level concession, not a per-model rate. It does not appear in the model-page prices above; those are the list rates a GET /logs call will report against a model, not a promotional rate.

Why the pages you already found are priced against a 2026 lineup that has moved on

The three pages that actually attempt video pricing (the ones ranking alongside the token tables) were all published in the first quarter of 2026, and their model lists show it:

Those are the dates and the model names printed on the pages, read on September 15, 2026. Both facts matter. Runway and Replicate publish their own rates on their own pricing pages (Runway, Replicate), so nothing stops a page from quoting them accurately; the issue is coverage, not accuracy. Between them, these three pages name 16 video models: 9 rows in TeamDay's per-second table, plus 4 names Renderful adds (Kling 1.6, Runway Gen-3, Hailuo MiniMax, Wan 2.1), plus 3 more that Crazyrouter adds (Veo 3, Luma Ray 2, Pika 2.2). The catalog measured above has 112, and seven of the names on the current list (seedance-2.0 and seedance-2.5, kling-v3, viduq3, wan3.0, hailuo-2.3, grok-imagine-video) appear on none of the three. One current name does cross over, and it is worth naming: TeamDay prices Veo 3.1 at $0.20/sec on a table whose own footnote reads July 2026, so that page has been refreshed since publication, while veo-3.1-lite, the cheaper Veo route in today's catalog, is still absent from all three.

That gap is the whole reason a per-second table is worth maintaining. Video model pricing is not a stable list you can publish once and leave; the models turn over faster than the articles about them do.

One name deserves a straight answer rather than a row in a table. Sora. It appears on most competing lists, and it is the video brand readers most often arrive looking for, but it is not in the 112-model catalog this page measured, so there is no first-party per-second number we can give you for it. Rather than estimate, we are saying so: for Sora's API rates, read OpenAI's own pricing page. For Google's Veo rates on Google's own platform, the comparable first-party source is Vertex AI's pricing page, and for MiniMax's Hailuo family it is the MiniMax platform.

How to check what you actually paid, and pull the price parameters yourself

A price table is a forecast. The invoice is the truth, and the two disagree for boring reasons: a parameter you did not realize you changed, a task routed to a different variant. Three things make the reconciliation mechanical instead of forensic.

1. The per-request cost log. Every call returns a log record you can read back at GET /api/v1/logs, and the response carries the model that actually ran and what it cost:

curl --request GET \
  --url 'https://api.modellix.ai/api/v1/logs?page=1&page_size=10' \
  --header 'Authorization: Bearer <token>'

The response shape, exactly as documented: a requests array whose entries carry task_id, status, the resolved model.provider and model.model_name, created_at, and cost. The model.model_name field is the one that settles arguments: it is what ran, not what you asked for. Read cost in the right unit: the reference documents it in sub-pennies, where 1 equals USD $0.0001, so divide by 10,000 for dollars. A row reading cost: 12 is $0.0012, and a cost of 0 means the task was not charged, which is how you check whether a failed generation billed. Field names and endpoint come from the Get Logs reference, read September 15, 2026.

The GET /api/v1/logs response documented on docs.modellix.ai, captured September 15, 2026. The cost field on each request is what makes a list-price table checkable against your own bill.

2. The price parameters, without guessing. The reason a table row is a range is a field in the request body: resolution, duration, task variant. You do not have to take a blog's word for which field it is: modellix-cli model get-schema returns any model's request and response schema, and it needs no API key:

modellix-cli model get-schema bytedance/seedance-2.0-t2v

Use the exact provider/model slug, as it appears in the table above. JSON is the default output; --output human summarizes the contract and --quiet prints just the inference URL. Pull the schema for two routes you are choosing between and diff them. That is a more reliable answer to "what actually moves this price" than any comparison post, including this one. The CLI reads the public schema endpoint described in Get schema, and its own usage is documented under Ways to use the CLI, both read September 15, 2026.

3. The input side, capped at seven days. Per-second billing describes the output. The inputs (a reference image, a first frame, a video to restyle) have to get to the model somehow, and the File API takes an upload without charging for it: POST /api/v1/media/files accepts an image, video or audio file as multipart form data and returns a file_id plus a url you pass straight into a prediction input field such as image_url. One constraint to plan around, quoted from the changelog: uploaded files are "retained for about 7 days by default." Build for that window rather than treating the upload as permanent storage (File API changelog entry, and the endpoint at Upload media file). Audio belongs in that sentence as an input format only; the models on this page generate video.

curl --request POST \
  --url https://api.modellix.ai/api/v1/media/files \
  --header 'Authorization: Bearer <token>' \
  --form 'file=@first-frame.png'

Two boundaries on this snapshot. Video generation is asynchronous: the call returns a task you retrieve rather than a finished file, and cost appears in the log after retrieval. Model paths no longer carry an /async suffix; the old form still works. These prices do not establish that Modellix is the cheapest host. At publication, we distributed the 112 routes listed here, including the most expensive one.

How to choose when the table will not crown a winner

A price table cannot pick for you, because the cheapest per-second rate and the right model are different questions. Narrow it with the reason you are shopping:

Three habits that survive contact with a real budget. Run the same prompt across two candidates for a week before committing: per-second rates are published, but how many usable clips you get per hundred attempts is not, and that ratio dominates the arithmetic. Set a low-balance alert so a batch job that silently switched to a pricier resolution surfaces in hours rather than at month end. And re-read the model page before you sign off on a forecast, because a rate captured in September is a September rate; this page says September 15, 2026 on every number for that reason.

The rate is published. The yield is not. Until you have run your own prompts through a route, you are comparing the half of the arithmetic that everybody else already knows.

Frequently Asked Questions About AI Video Model API Pricing

How much does an AI video model API cost per second? Across the 112 video models on sale at Modellix on September 15, 2026, per-second list prices ran from $0.0050 to $1.4400. The median lowest price was $0.0710/sec, which is about $0.36 for a five-second clip. The floor belongs to a digital-human route (vidu/viduq2-pro-digital-human) and the ceiling to a video-editing route (alibaba/happyhorse-1.0-video-edit), so the honest answer to "what does a video second cost" depends on which task you are buying.

Why do some video models list a price range instead of one number? Because the price is a function of parameters you set, most often resolution, sometimes duration or task variant. 88 of the 112 routes publish a range for that reason. bytedance/seedance-2.0-t2v runs $0.0700/sec at the low end and $0.7800/sec at the top, an 11.1× spread on the same model. To find which field controls it for a specific route, pull the model's request schema rather than trusting a table.

How much does Kling API access cost? Kling's cheapest route in the catalog is kling/kling-avatar at $0.0448~$0.0896/sec, and the text-to-video routes start at $0.0672/sec (kling/kling-v3-t2v, up to $0.3360/sec) with kling/kling-v3-turbo-t2v from $0.0896/sec. Kling accounts for 12 of the 112 video models, all measured September 15, 2026. For the full single-vendor breakdown, see our Kling 3.0 cost breakdown.

How much does Veo API pricing come to per second? In this catalog, Google's Veo routes start at $0.0450/sec for google/veo-3.1-lite-t2v (ceiling $0.0720/sec) and reach $0.3600~$0.5400/sec for google/veo-3.1-t2v. google/veo-3.1-fast-t2v sits between them at $0.0900~$0.2700/sec. Google's own platform prices the same family on its Vertex AI pricing page; the higher-tier Veo numbers are also broken out in our Veo 3.1 price page.

How much does Seedance API pricing run? ByteDance's Seedance routes are among the widest ranges in the catalog: bytedance/seedance-2.0-mini-t2v starts at $0.0400/sec, bytedance/seedance-2.0-t2v spans $0.0700~$0.7800/sec, and bytedance/seedance-2.5-t2v spans $0.1030~$0.5690/sec. The mini and video-to-video variants are the low-cost entry points; the per-generation price ladder is in our Seedance per-second price page.

Is there a free tier for video model APIs? No. Modellix's registration credit was discontinued on August 19, 2026, and new accounts do not receive an automatic credit. Modellix does not charge for File API uploads, but that is an input-transfer exception, not free generation. Treat any "free video API credits" claim on a third-party list as needing a date attached.

What is the cheapest way to compare video model APIs fairly? Normalize to cost per usable clip, not cost per second. Take your real prompt volume, run two candidate routes at different resolutions for a week, count how many outputs you would actually ship, and divide the total spend by that number. 24 of the 112 routes publish a flat per-second price, which makes them easier to forecast; the other 88 will move with a parameter, so the comparison only holds if you hold that parameter fixed across both sides.