Finding the cheapest video generation API for what you actually run
Updated 2026-10-02
"Cheapest video generation API" has no single answer, because the cheapest option depends on the model you need, the resolution you ship, how often you retry and what you do when a provider is down. A list sorted by headline per-second price answers a different question than the one on your invoice. This page is a method for getting to your answer.
Step 1: fix the thing you are buying
Price is only comparable once three things are held constant:
- The model. Different checkpoints are different products. "Cheapest" between Kling, Veo and Seedance is a model choice, not a pricing choice, and belongs after you have decided which outputs are acceptable.
- The resolution tier. Video is priced per second per tier, so a 480p rate against a 1080p rate is not a comparison. Pick the tier you actually ship.
- The clip length and mode. Text-to-video, image-to-video and reference modes can price differently. Some models also bill per clip or per request instead of per second, which the per-second tables do not capture.
The same model is often sold by several hosts at different prices. That is where most of the savings are, and it is why the table below lists the cheapest and priciest host for each model at one named tier.
| Model | Cheapest host | Priciest host | Cheapest is | Hosts |
|---|---|---|---|---|
| bytedance/seedance-2.5 (480p) | OpenSand $0.0525 / second | Fal-US $0.2646 / second | 80% lower | 9 |
| bytedance/seedance-2.0 (2160p) | MachGen $0.59 / second | Fal $1.5552 / second | 62% lower | 9 |
| alibaba/wan-3.0 (480p) | Replicate $0.025 / second | Alibaba $0.05 / second | 50% lower | 10 |
| google/gemini-omni-flash | Google $0.1 / second | Fal $0.13 / second | 23% lower | 4 |
| minimax/h3 (768p) | MachGen $0.04 / second | WaveSpeedAI-resell $0.1 / second | 60% lower | 14 |
| alibaba/happyhorse-1.1 (720p) | Pika $0.098 / second | Alibaba $0.14 / second | 30% lower | 4 |
| bytedance/seedance-2.0-fast (480p) | Atlas Cloud $0.027 / second | Fal $0.2419 / second | 89% lower | 9 |
| bytedance/seedance-2.0-mini (480p) | OpenSand $0.0104 / second | Fal $0.0721 / second | 86% lower | 8 |
| seedance-2-mini-unrestricted (480p) | OpenSand $0.0114 / second | SandBase $0.0721 / second | 84% lower | 3 |
| kling-o3 (720p) | SandBase $0.0588 / second | Tencent TokenHub $0.084 / second | 30% lower | 4 |
| minimax/h3-max (480p) | SandBase $0.01 / second | MiniMax $0.05 / second | 80% lower | 5 |
| kling-v3 (2160p) | SandBase $0.294 / second | Tencent TokenHub $0.42 / second | 30% lower | 8 |
Per second, before VideoRouter's 2% platform fee. For tiered models each row compares the resolution tier with the widest host-to-host gap. Built 2026-10-02 from the live catalog.
Step 2: compare hosts at your tier
Within one model and tier, the spread between the cheapest and priciest host is often larger than the gap you would get from switching models. Check the model's page for the full host-by-tier grid rather than the summary row, because the cheapest host at one tier can differ from the cheapest at another. The pricing page carries the live comparison and this explainer covers the billing mechanics.
Step 3: add the multipliers the rate card leaves out
| Factor | How it changes cost | What to measure |
|---|---|---|
| Retries | You pay for every successful generation, including the ones you discard | Attempts per accepted clip, from a pilot on your own prompts |
| Failed jobs | On VideoRouter, jobs that fail upstream are not billed; check this for every provider you compare | Billing policy for failures, not just price |
| Platform or reseller fee | VideoRouter adds a flat 2% on top of the host price for image and video; others differ | Whether the quoted price already includes the fee |
| Minimum spend and tiers | Commitments and minimum top-ups raise effective cost at low volume | Minimums, expiry of credits, volume tiers |
| Duplicate attempts | Timeout hedging can bill two attempts if both finish | Whether you enabled it, and for which jobs |
| Non-per-second pricing | Some models charge a flat amount per clip or per request regardless of duration | The unit on the model page |
| Storage and egress | Keeping and serving finished files is your cost, not the API's | Retention policy and delivery bandwidth |
Step 4: price reliability, not just the rate
A cheap host that fails often is expensive in engineering time and user-visible errors. Two things help:
- Automatic failover. With no provider specified, requests go to the cheapest healthy host and walk down the list if a submission is rejected, so a single provider outage does not stop you. Pinning one host gives that up, so pin only for a reason.
- Performance-aware ordering. You can rank hosts by measured reliability or latency instead of price when consistency matters more, accepting a higher rate for it.
Prefer a stable host over the absolute floor when a failed generation costs you more than the savings. Whether that is true depends on your product, so decide it deliberately.
Step 5: run a small pilot
List prices cannot tell you your retry factor. Run 30 to 50 representative prompts through your shortlist at the tier you ship. Record cost per request, attempts per accepted clip, failure rate and time to completion. Then compute cost per accepted second, which is the metric that matches your invoice:
cost_per_accepted_second = (total_billed_seconds * rate * (1 + fee)) / accepted_secondsA host with a lower rate and a worse acceptance ratio can lose to a pricier host that wastes less. If your clips go through a draft-then-final workflow, run the pilot on both stages separately.
Common mistakes
- Comparing across tiers. A "from" price is usually the lowest tier. Always compare at the tier you will ship.
- Ignoring the fee line. Check whether a quoted price is the provider's own rate or includes a platform fee, then compare the same way for every option.
- Locking in on one host. Rates move. Revisit the comparison on a schedule rather than hard-coding a host forever.
- Optimizing the wrong lever. If retries dominate your spend, a cheaper host helps less than better prompts, shorter drafts, or caching accepted clips.
To turn the result into a monthly figure, use the budget method, or see the cost page for live clip costs. You can try the cheapest-healthy-host routing yourself with a key from videorouter.sh/signup, and the quickstart has the code.
Frequently asked questions
What is the cheapest AI video generation API?
It depends on the model, the resolution tier and your retry rate. Compare hosts for the same model at the same tier, then add retries, fees and reliability.
Is the cheapest host per second always the cheapest overall?
No. A lower rate with more failures or discarded attempts can cost more per accepted clip than a pricier, steadier host. Measure cost per accepted second in a pilot.
Do failed generations cost money?
On VideoRouter, jobs that fail upstream are not billed. Policies differ between providers, so check each one you compare.
How big is the platform fee?
VideoRouter adds a flat 2% on top of the host price for image and video generation, with no tiers and no minimum spend.
Keep reading
- How AI Video API Pricing Works — Per-Second, Per-Tier, Per-Host
- Video API Pricing by Resolution: 480p to 4K Per-Second Costs
- How to Estimate Your AI Video API Budget, Step by Step
- AI Video Generation Cost Per Minute: How to Calculate It
VideoRouter puts it next to dozens of other video and image models behind one API key, so you can compare providers, prices and fail over automatically. Compare providers on VideoRouter →