Video generation is the one part of the API market where nobody publishes an honest unit price. Every provider quotes "per second", but what one second actually costs depends on resolution, duration rounding, whether frames are billed as input, and how the task is settled. So we did the boring thing: we called nine video models with an identical 4-second prompt, then read the billing delta from the upstream account before and after each call.
All numbers below are paid, measured, USD per second of generated video. Prompt: a calm sea at sunrise, duration 4s, single call each, CNY→USD at 6.8 (we settled in CNY).
Measured upstream cost
| Model | Upstream cost (USD/sec) | 4s call (USD) | Wall time |
|---|---|---|---|
| grok-imagine-video | 0.0132 | 0.053 | 67s |
| veo-3.1-fast | 0.0199 | 0.080 | 55s |
| grok-imagine-video-1.5 | 0.0212 | 0.085 | 90s |
| gemini-omni-flash-1.1 | 0.0265 | 0.106 | 61s |
| gemini-omni-flash-1.2 | 0.0265 | 0.106 | 61s |
| veo-3.1 | 0.0331 | 0.132 | 48s |
| doubao-seedance-2.0-fast | 0.0802 | 0.321 | 139s |
| doubao-seedance-2.0 / 2.5-flash | 0.0997 | 0.399 | 134–158s |
| wan3.0-video-prime | 0.2118 | 0.847 | 140s |
Three findings that surprised us:
1. The spread is 16×. grok-imagine-video and wan3.0-video-prime are the same task at the same length, and one costs 16 times the other. If you are picking a video model by "which one looks newer", you are picking your bill at random.
2. Wall time does not track price. The cheapest model finished in 67s; the second cheapest took 55s; the most expensive took 140s. Cheap and slow is a real trade-off, but "expensive" bought us nothing in latency either.
3. Some models are simply not there. kling-3.0, kling-3.0-omni and kling-3.0-motion-control returned 502/400 on submit; seedance-2.0* and gemini-omni-flash (non-1.1/1.2) returned model_not_found with no available channel. Five of sixteen candidates did not produce a single frame. We removed them from our catalogue rather than letting them fail after your credit is held.
What it means if you are buying
A 4-second clip at 0.0132/sec costs 5 cents. The same clip at 0.2118/sec costs 85 cents. Run 100 clips a month and the choice is $5 versus $85 — for outputs most people cannot tell apart without a side-by-side.
Two practical rules we now apply internally:
- Price per second × duration is the only comparable number. Per-call pricing hides the duration multiplier; per-resolution pricing hides the frame billing.
-
Test the settle, not the submit. A task that returns
queuedproves nothing. We only treat a model as available after the file URL comes back and the billing delta lands.
How we call them
Our gateway exposes the same nine models behind one endpoint, so switching is a one-line change:
curl https://tapixai.com/v1/videos \
-H "Authorization: Bearer $TAPIX_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"veo-3.1-fast","prompt":"a calm sea at sunrise","duration":4}'
Then poll the returned id:
curl https://tapixai.com/v1/videos/$TASK_ID -H "Authorization: Bearer $TAPIX_KEY"
You get status: queued → completed and a file URL. Billing is per second of generated video, charged only when the task completes; a failed task releases the hold.
We publish these numbers because we would rather you trust a bill you can reproduce than a comparison table you cannot. If a cheaper model appears upstream, it shows up in our catalogue the next morning.
New accounts start with $1 of credit — enough for roughly 5–50 of the clips above depending on which model you try. No card required.
Top comments (0)