AI Video Generation API Pricing in 2026, Model by Model

Ten seconds of AI video can cost seventy-five cents or four dollars, and most of that gap comes from the model name on the invoice. HackerNoon's July 2026 breakdown put a 10-second Kling 3.0 clip at about $0.75 and the same length on Veo 3.1 Standard at about $4.00. Across the whole market, invideo's August 2026 survey of official rates found per-second prices running from $0.02 to $0.70, a spread of roughly 35x.

If you are wiring video generation into a product, a client pipeline or an agent, the sticker price per second is only the first input. The billing unit, the resolution tier, whether audio is bundled, and how many attempts you burn per usable shot decide what you actually pay. This guide covers each of those, then gives a model-by-model rate list and three worked batch budgets you can adapt.

The three ways video APIs bill you

Almost every provider uses one of three units. Knowing which one you are on tells you how to forecast, and where the surprises hide.

Per second of output

This is the dominant model. You pay a rate multiplied by the length of the clip you asked for. Google's Gemini API pricing page lists Veo 3.1 this way, and Replicate bills its Wan 2.1 image-to-video models per second of output video ($0.09/s at 480p, $0.25/s at 720p). Forecasting is simple: seconds times rate. The catch is that every tier change (resolution, audio, a "Pro" variant) is a different rate, so the same script can price very differently depending on one parameter.

Per clip or per unit

Some models charge a flat price per generated video regardless of length. fal.ai notes that video models are billed by output unit, per second or per video depending on the model, and lists Ovi at $0.2 per video. Per-clip pricing favours long clips and penalises short cuts. Replicate bills most of its other models by run time on the underlying hardware, which is harder to predict before you run a sample.

Credits

Credits are a currency layer on top of one of the units above. Runway's API pricing is the clean version: one credit costs $0.01, and each model consumes a fixed number of credits per second (gen4.5 uses 12 credits/s, so $0.12/s). If you are weighing a subscription against metered billing, the credits versus pay-per-call comparison works through when each one comes out cheaper.

A practical rule: always convert to dollars per second of usable output before comparing anything. A credit price, a per-video price and a run-time price are not comparable until you do.

What actually moves the price

Coverr's 2026 pricing guide groups the drivers into resolution and audio, model choice, billing structure and infrastructure overhead. The primary sources show how large each lever is.

Model tier

Within a single family, the tier is the biggest lever. On the Gemini API, Veo 3.1 Standard costs $0.40/s at 720p and 1080p, Veo 3.1 Fast costs $0.10/s at 720p, and Veo 3.1 Lite costs $0.05/s at 720p. That is an 8x spread inside one model name. Runway shows the same pattern with gen4.5 at 12 credits/s against gen4_turbo at 5 credits/s.

Resolution

Resolution steps are rarely linear. Veo 3.1 Fast on the Gemini API goes from $0.10/s at 720p to $0.12/s at 1080p, then $0.30/s at 4K. Runway's wan3 goes 5, 10 and 20 credits/s across 480p, 720p and 1080p, doubling at each step. For most social and product work, the useful question is whether the extra pixels survive the platform's compression.

Audio

Native audio is sometimes bundled and sometimes billed as a separate tier. Google states that Veo pricing includes video with audio by default. Runway lists its Veo 3.1 entries with audio (40 credits/s for veo3.1, 15 credits/s for veo3.1_fast). Elsewhere the silent and audio versions carry different rates, so check which one you are calling. If you plan to lay a voiceover and music over the clip anyway, paying for generated audio is money you throw away in the edit.

Add-ons and references

Runway notes that some models add per-reference charges on top of the per-second rate. Image references, character references and extensions can each carry their own line item. Read the model's page, not just the headline rate.

Failed renders

Google says you are only charged if the video is successfully generated. Not every provider documents this, and on a batch of hundreds of jobs the difference between "failed jobs are free" and "failed jobs are billed" shows up clearly on the invoice.

Market rates at the time of writing

Prices change often, so treat every number below as dated to its source. Primary pages are marked as such. Where a figure only exists in a secondary source, it is labelled that way.

First-party and platform rates (primary sources)

Veo 3.1 on the Gemini API: Standard $0.40/s (720p and 1080p), $0.60/s at 4K. Fast $0.10/s at 720p, $0.12/s at 1080p, $0.30/s at 4K. Lite $0.05/s at 720p, $0.08/s at 1080p, no 4K. Source: Google AI for Developers.
Runway API: gen4.5 $0.12/s, gen4_turbo $0.05/s, seedance2 $0.36/s at 480p/720p and $0.40/s at 1080p, seedance2_fast $0.29/s, seedance2_mini $0.16/s, veo3.1_fast with audio $0.15/s, veo3.1 with audio $0.40/s. Source: Runway.
fal.ai: Wan 2.5 $0.05/s, Kling 2.5 Turbo Pro $0.07/s, Veo 3 $0.40/s, Ovi $0.2 per video. Source: fal.ai.
Replicate: wan-2.1-i2v-480p $0.09/s, wan-2.1-i2v-720p $0.25/s of output. Source: Replicate.

Rates reported by secondary sources (verify before you budget)

Sora 2 API: reported at $0.10/s at the base tier and up to $0.70/s for Sora 2 Pro by invideo and HackerNoon. OpenAI's own pricing was not confirmed for this article, so treat these as indicative.
Kling 3.0: about $0.075/s per HackerNoon. invideo reports Kling 3.0 Turbo at about $0.11/s, 720p only.
Seedance 2.x direct: billed per million tokens, which invideo approximates to about $0.14/s.
Grok Video: $0.05/s for 1.0 and $0.08/s for 1.5, per invideo.
Runway access: Coverr reports that Runway API access requires a Max plan at $76+ per month. That is a secondary claim; check Runway's current terms.

Two cautions. Version numbers differ across hosts (fal lists Wan 2.5 and Kling 2.5 Turbo Pro, Replicate lists Wan 2.1), and newer versions price differently. And the same model can cost different amounts on different hosts. Veo 3.1 Fast with audio is $0.10/s at 720p on the Gemini API and 15 credits/s ($0.15/s) on Runway.

The aggregator rate list: one key, many models

For comparison, here is the per-call grid on Aitachyon at the time of writing. Each entry links to the model's own price page, and the full machine-readable grid is published at /api/pricing so a script can read it instead of a human copying numbers.

Video

Seedance 2.5 (ByteDance): from $0.20/s at 480p, $0.44/s at 720p.
Kling v3 (Kuaishou): $0.16/s silent, $0.32/s with native audio.
Veo 3.1 Fast (Google): from $0.19/s.
Wan 2.7 (Alibaba): $0.19/s.
Hailuo 02 (MiniMax): $0.085/s.

Images and voice (for the stills and narration that go around the video)

Seedream 4.0: $0.057 per image.
FLUX.2 [pro]: $0.057 to $0.086 per image.
Nano Banana: $0.13 per image.
gpt-image-2: $0.10 to $0.40 per image.
ElevenLabs voiceover: $0.043 to $0.086 per 450 characters.

Read against the primary sources, the trade-off is plain. Calling Veo 3.1 Fast directly on the Gemini API is cheaper per second than calling it through an aggregator. What an aggregator sells is the rest: one key and one balance across vendors, one invoice line per job, and the ability to route each shot to a different model without opening five accounts. Coverr's comparison of 100 eight-second clips a month puts Veo 3.1 direct at about $160 to $320 and aggregators such as fal.ai and Replicate at $40 to $200, and most of that gap comes from routing volume to cheaper models. For a deeper look at two of these models, see Kling v3 against Seedance 2.5 and the Veo 3.1 Fast API breakdown.

Cost per usable clip: the formula that matters

The number on a pricing page is the cost of an attempt. The number you should budget is the cost of a clip you keep. HackerNoon reports that professional work commonly runs three to five generation passes per hero shot, and recommends capping regeneration at three attempts before rethinking the approach. That puts the real cost of a hero shot at three to five times the sticker price.

Use this formula for every budget:

Attempt cost = rate per second x seconds requested (or the flat per-clip price).
Add-ons = reference, audio or upscale charges for that call, if any.
Attempts per keeper = your measured average. Start with 3 for hero shots, following HackerNoon, and replace it with your own number after the first batch.
Failed-render cost = zero if the provider refunds or does not bill failures, otherwise add the failure rate.
Cost per usable clip = (attempt cost + add-ons) x attempts per keeper + failed-render cost.

A worked example: a 10-second Kling v3 silent clip at $0.16/s is $1.60 per attempt. At three attempts the keeper costs $4.80. At five it costs $8.00. The same shot on Hailuo 02 at $0.085/s is $0.85 per attempt and $2.55 at three attempts. When the attempt count dominates, a cheaper model you can iterate on quickly often beats a pricier one you have to reroll less often, and the only way to know is to log attempts per keeper by model.

To cut attempt cost, test prompts at the shortest length the model allows, lock the composition, then render the full length once.

Three worked batch budgets

These use the Aitachyon rates listed above, at the time of writing. The retry assumptions are stated so you can swap in your own.

Batch 1: 20 video hooks for a creative test

Twenty 5-second opening shots, each tested as the first beat of a paid social video. Hooks are low-stakes and short, so assume 2 attempts per keeper.

Hailuo 02: 20 x 5 s x $0.085 = $8.50 per pass, $17.00 at 2 attempts.
Kling v3 silent: 20 x 5 s x $0.16 = $16.00 per pass, $32.00 at 2 attempts.
Seedance 2.5 at 720p: 20 x 5 s x $0.44 = $44.00 per pass, $88.00 at 2 attempts.

The model choice changes the batch cost by about 5x for the same brief. For a hook test, where the point is to learn which opener earns attention, the cheap model is usually the right call. The hook ideas themselves matter more, and these hook templates give you twenty prompts to start from.

Batch 2: 50 product shots

Fifty still images for a catalogue refresh or static creative, one attempt each plus a 50% reroll allowance (75 generations total).

Seedream 4.0: 75 x $0.057 = $4.28.
FLUX.2 [pro]: 75 x $0.057 to $0.086 = $4.28 to $6.45.
Nano Banana: 75 x $0.13 = $9.75.
gpt-image-2: 75 x $0.10 to $0.40 = $7.50 to $30.00.

Stills are an order of magnitude cheaper than video. If some of those 50 shots then get animated, animate only the winners. Five selected shots turned into 5-second Kling v3 silent clips at three attempts each add 5 x 5 x $0.16 x 3 = $12.00.

Batch 3: a week of shorts

Seven shorts, each built from four 5-second clips (20 seconds of video), a voiceover script of up to 900 characters, and one cover image. Assume 2 attempts per clip.

Video on Kling v3 silent: 7 x 20 s x $0.16 x 2 = $44.80.
Voiceover: 900 characters is two 450-character units, so $0.086 to $0.172 per short, $0.60 to $1.20 for the week.
Covers on Seedream 4.0: 7 x $0.057 = $0.40.
Week total: about $45.80 to $46.40.

Swap Kling silent for Kling with native audio ($0.32/s) and the video line doubles to $89.60, which is why pairing silent video with a separate voiceover is usually the cheaper stack when narration carries the message. The ElevenLabs voiceover cost breakdown goes further on narration pricing.

Routing rules: which model for which shot

With per-call pricing, choosing a model is a per-shot decision rather than a subscription you commit to. These rules use only the price and spec differences documented above.

Exploration and hook tests: use the cheapest model that produces an acceptable composition. Hailuo 02 at $0.085/s is the low end of this grid.
Narrated content: use silent video and add a separate voiceover. Kling v3 silent at $0.16/s plus ElevenLabs costs less than Kling with audio at $0.32/s for the same seconds.
Shots that need synced native sound (dialogue, diegetic noise): pay for the audio tier, and budget it as double the silent rate where the model has one.
Resolution: start at the lowest tier that survives your destination's compression. Seedance 2.5 costs $0.20/s at 480p and $0.44/s at 720p; move up only for a keeper.
Hero shots: spend on the premium model only after the prompt is locked on a cheap one, and cap attempts at three as HackerNoon recommends.
Stills first: if an image can do the job, it costs cents. When a static beats a video covers where that holds.

Automating generation without losing control of spend

Batch budgets stay honest when generation runs from code, because code logs every call and a web UI invites rerolls nobody counts. Developers and small teams producing at volume drive generation from a script, or from an agent such as Claude Code or Cursor through an MCP server: a prompt or job file in, files with their per-call cost out.

A pre-flight checklist for any automated batch

Pin the model and the tier explicitly in the call. Never rely on a default resolution or audio setting.
Read the price grid programmatically before the run and compute the batch ceiling: jobs x seconds x rate x expected attempts.
Set a hard attempt cap per shot in code (three is a sane default for hero shots).
Use a dedicated API key per project or per client so spend is separable.
Confirm how failed renders are billed before you run hundreds of jobs.
Store a stable reference for every output and the cost of the call that produced it, so you can compute cost per keeper afterwards.
Review attempts per keeper by model after each batch and update your routing rules.

Agents make this faster and also riskier: an agent will happily reroll a shot ten times if nothing stops it. Per-key spend alerts and a revocable key are the guardrails. For a working agent setup, see generating video from Claude Code with an MCP server, and for the API and MCP quick start, see the developer docs.

FAQ

How much does an AI video generation API cost per second?

Official rates at the time of writing run from about $0.02 to $0.70 per second, according to invideo's August 2026 survey. On Google's own API, Veo 3.1 ranges from $0.05/s (Lite, 720p) to $0.60/s (Standard, 4K).

Is it cheaper to use a video API directly or through an aggregator?

For a single model at volume, going direct can be cheaper per second, as with Veo 3.1 Fast on the Gemini API. Aggregators win when you route different shots to different models, since cheaper models on a shared balance pull the average down. Coverr's example of 100 eight-second clips per month puts Veo 3.1 direct at $160 to $320 against $40 to $200 on aggregators.

Do you pay for failed AI video generations?

It depends on the provider. Google states that Veo is only charged when the video is successfully generated. Check each provider's terms, because on large batches failed-job billing adds up.

How many attempts should I budget per AI video clip?

HackerNoon reports three to five passes per hero shot in professional work, and suggests capping at three before changing approach. For low-stakes clips such as hook tests, measure your own rate after the first batch and budget from that.

Sources

If you want to run the batches above without opening an account per vendor, Aitachyon puts these video, image and voice models behind one prepaid balance that never expires, callable from the web studio, a plain HTTP API or a hosted MCP server. Every job is itemised with its model and cost, failed renders are refunded automatically, and each key has its own spend alert and one-click revoke. The prices are on every model's page.