If you have priced out AI video generation lately, you already know the headline numbers rarely tell the full story. One provider quotes a price per second, another sells credits, a third bundles everything into a flat per-clip fee, and none of them make it easy to compare against the others.
This guide breaks down what the major AI video API providers actually charge in 2026, converts everyone's pricing into a common unit where possible, and walks through the extra costs that tend to show up after you have already committed to a vendor. Prices change often in this space, so treat the numbers below as a snapshot and always confirm current rates on the provider's own pricing page before you budget.
How AI Video API Pricing Actually Works
Video generation pricing tends to follow one of three models.
Per-second pricing charges you based on the length of the clip you generate, sometimes with a different rate for each resolution or quality tier. This is the most transparent model because the math is simple: rate times duration equals cost.
Per-clip pricing charges a flat fee for a fixed-length generation, usually in preset buckets like 5 seconds or 10 seconds rather than any duration you want. Providers that use this model often still vary the price by resolution, mode, or whether audio is included.
Credit-based pricing sits on top of either of the above. You buy a block of credits up front, and each generation deducts a certain number of credits depending on model, resolution, duration, and features like audio or 4K output. The credit itself has a dollar value, but that value can shift depending on which plan you are on, since higher tiers usually give you a lower effective cost per credit.
Resolution and duration are the two levers that move your bill the most. Going from 720p to 1080p can roughly double the cost on some models, and adding audio generation frequently adds another 30 to 100 percent on top of the silent price.
Pricing by Provider
Runway (Gen-4 Turbo and Gen-4.5)
Runway's API bills in credits at a fixed rate of $0.01 per credit. Gen-4 Turbo runs 5 credits per second, which works out to $0.05 per second, or about $0.50 for a 10-second clip. Gen-4.5, Runway's flagship model, runs 12 credits per second ($0.12/sec), so a 10-second clip costs roughly $1.20. Runway also charges per-second rates for its Veo 3 and Veo 3.1 integrations and for newer additions like Seedance 2.0, so the API effectively acts as a marketplace with Runway's own margin baked into each rate.
Kling (3.0)
Kling's official API sells prepaid resource-unit packages rather than a simple pay-as-you-go rate, starting around $9.80 for a small trial package and scaling into much larger blocks for production use. Per the published Kling 3.0 model guide, generation without audio runs roughly 6 to 8 credits per second depending on resolution (720p vs 1080p), climbing to 9 to 12 credits per second with native audio, plus a small add-on for voice control. A useful detail: failed API generations do not consume credits, which is not always true on the consumer app.
Hailuo (2.3)
MiniMax's Hailuo API is billed per second and scales with resolution. Third-party reseller pricing puts 512p generation around $0.01 per second, 768p around $0.04 per second, and 1080p around $0.08 per second, though MiniMax also sells prepaid monthly Token Plans starting at $10/month for teams that want predictable billing instead of metered usage. Hailuo 2.3 Fast is the discount tier, trading some quality for a noticeably lower per-second rate at the same resolution.
Seedance (2.0)
ByteDance's official Seedance pricing (via Volcengine) is quoted per million tokens rather than per second directly: 46 CNY per million tokens for straight generation and 28 CNY per million tokens for video-input/editing mode, which works out to roughly 1 CNY (about $0.14) per second on a 15-second clip at published rates. Several third-party API resellers undercut that significantly, offering 720p generation from around $0.045 to $0.05 per second, so shopping around matters more with Seedance than with most other models.
Luma Dream Machine
Luma prices everything in credits, purchased through monthly plans: Plus at $30/month for 10,000 credits, Pro at $90/month for 40,000 credits, and Ultra at $300/month for 150,000 credits, which works out to roughly $0.002 to $0.003 per credit depending on the plan (higher tiers get a slightly better rate). Video cost scales steeply with resolution: a 720p generation on Luma's Ray model runs about 20 credits per second (roughly $0.06/sec on the Plus plan), while 1080p jumps to 80 credits per second (about $0.24/sec). Luma's credits cover both the web app and API access, and unused credits from top-up packs roll over for 12 months, unlike the monthly subscription allotment.
Sora API
OpenAI's Sora 2 API prices by resolution and tier: Sora 2 standard runs $0.10 per second at 720p (or $0.05/sec on the batch tier), while Sora 2 Pro scales from $0.30 per second at 720p up to $0.70 per second at 1080p on the standard tier, roughly half that on batch. The bigger story here isn't the price, it's the calendar: OpenAI has confirmed the Sora 2 API and its video generation model snapshots will be removed on September 24, 2026, with no successor API announced yet. If you have production traffic on Sora today, budget time now to test a replacement model rather than migrating under deadline pressure later this year.
Hidden Costs That Inflate Your Bill
The headline per-second or per-credit rate is rarely the whole story. Audio generation is usually an add-on that can raise the effective cost by 30 to 100 percent over a silent clip on the same model. Upscaling a generation to 4K or a higher frame rate is typically billed separately, on top of the original generation cost. Failed generations are the quiet budget killer: some providers refund credits automatically when a job fails, others do not, and it is worth checking this before you build retry logic that could otherwise burn through your balance. Storage and CDN egress fees can also creep in if a provider only hosts your output for a limited window (Apiframe, for example, keeps generated files on its CDN for 90 days) and you need to pull large volumes of video out for archival.
How to Estimate Your Monthly Spend
The formula is simple: clips per month multiplied by average cost per clip equals monthly spend. The hard part is being honest about the "average cost per clip" number, since it should include your actual mix of resolutions, durations, and retry rate, not just the cheapest tier you found on a pricing page.
A rough worked example across three volume tiers, using a mid-range per-second rate around $0.06 (roughly Kling 3.0 standard mode or Luma's 720p Ray tier):
A prototype building 20 test clips a month at 5 seconds each spends about $6 a month, which is well within most free trial credit allocations.
A growing app generating 500 clips a month at an average of 6 seconds spends around $180 a month, assuming a low retry rate and no premium resolution.
A production app at scale, generating 10,000 clips a month at 8 seconds average with a mix of standard and pro modes, can land anywhere from $3,000 to $12,000 a month depending on how much of that volume needs higher resolution or audio.
Those numbers shift quickly once you add audio generation or 4K output, so it is worth running the math on your actual expected mix before committing to a provider or a plan tier.
Cutting Video Generation Costs Without Sacrificing Quality
The single biggest lever is matching the model to the job instead of defaulting to the most capable one. A social media teaser clip rarely needs 4K or a flagship model; a fast, cheaper tier at 720p is often visually indistinguishable once it is compressed for a feed anyway.
Caching and reusing clips helps more than most teams expect, especially for evergreen content like product demos or onboarding videos that do not need to be regenerated every time someone views them.
Batching requests where a provider offers a discounted batch tier (Sora's batch pricing runs roughly half the standard rate) is worth building into your pipeline if your use case can tolerate slightly slower turnaround.
Perhaps the most practical lever if you are working across more than one model: using a single unified API pricing structure instead of juggling multiple vendor contracts, minimums, and credit systems. Apiframe gives you one API key and one credit balance across Kling, Hailuo, Seedance, Runway, Luma, Veo, and more, so you can route each job to whichever model is cheapest or best suited for it without negotiating separate accounts. You can see exact per-model, per-second credit costs on Apiframe's pricing page.
FAQ
What's the cheapest AI video API right now?
It depends on resolution and whether you need audio, but Seedance and Hailuo's faster tiers, along with Kling 3.0's standard mode, tend to land at the lower end of the per-second range compared to flagship models like Sora 2 Pro or Veo 3 with audio enabled.
Per-second vs. per-clip, which is actually cheaper?
Neither model is inherently cheaper, it depends on the underlying rate. Per-second pricing is easier to estimate precisely for variable-length content, while per-clip pricing can be a better deal if you consistently need durations that fall right at a provider's preset bucket.
Is there a free tier worth prototyping on?
Most providers offer some form of free trial credits, typically enough for a handful of test generations rather than sustained development. See our companion guide on free AI video generation APIs for a full breakdown of what each provider actually gives you before you have to pay.