AI video API benchmark (November 2026): render time, success rate and price per second for every model
How long does an AI video API take, how often does it deliver, and what does a second of video cost? We measured 68,309 real video jobs that customers sent through the Revana API across 24 models, plus 148,607 image jobs. Here are the numbers, with charts and a CSV you can download and cite.
- 34 smedian render, MiniMax H3 Max (768p), the fastest video model
- 99.1%of 48,947 Seedance 2.5 jobs delivered, the most reliable model at volume
- $0.03a second at 720p for Grok Imagine Video 1.5 Preview, the lowest 720p list price
- 72%of customer video jobs went to Seedance 2.5
Jawaban singkatnya
- Fastest: MiniMax H3 Max (34 s), MiniMax H3 (64 s), Grok Imagine Video 1.5 Preview (67 s), Gemini Omni 1.2 Flash (67 s), Gemini Omni (75 s). Median time from request to finished file, at each model's most used resolution.
- Slowest per second of video: the Seedance models. Seedance 2.5 took a median 358 s at 720p for a median 9 second clip, about 41 seconds of waiting for each second of video. MiniMax H3 Max needed 5.
- Most reliable: Seedance 2.5 delivered 99.1% of 48,947 jobs. Kling 2.5 Turbo Pro, Kling 4.0 Flash, Hailuo 02 Standard, Wan 2.7 delivered every job, on smaller volumes.
- Least reliable: Gemini Omni 1.1 Flash (83.4% over the period, 92.1% in the last 7 days) and MiniMax H3 (84.8%, but 124 of 124 in the last 7 days).
- Cheapest per second: Hailuo 02 at 512p ($0.02), then Grok Imagine Video 1.5 Preview at 720p ($0.03). Prices are Revana's list prices; see the pricing page for every rate.
Render time per model
Render time here is the wall-clock time from the moment the API accepted a request to the moment the finished file was stored and ready to download. It includes queueing and, when a route failed, the retry on the next route. So it is what a developer waiting on the API actually experiences, not a lab number.
Clip length matters: a 15 second clip takes longer than a 5 second one on the same model. To compare fairly, we also divided each job's render time by the length of the video it produced:
Two things stand out. First, the gap between the median and p90 is large on some models: MiniMax H3 had a median of 64 s but a p90 of 1285 s, because a few jobs waited a long time for capacity. If your product has a timeout, plan for the p90, not the median. Second, higher resolution is not always slower: for Seedance 2.5, 1080p jobs had a shorter median (237 s) than 720p (358 s) because people asked for shorter clips at 1080p (median 6 s against 9 s).
Render time by resolution
| Model | Resolusi | Delivered jobs | p50 | p90 | Median clip |
|---|---|---|---|---|---|
| Seedance 2.5 | 480p | 19,455 | 324 s | 556 s | 11 s |
| Seedance 2.5 | 720p | 28,322 | 358 s | 635 s | 9 s |
| Seedance 2.5 | 1080p | 723 | 237 s | 426 s | 6 s |
| Gemini Omni 1.1 Flash | 720p | 3,357 | 89 s | 160 s | 6 s |
| Gemini Omni 1.1 Flash | 1080p | 1,284 | 148 s | 224 s | 6 s |
| Gemini Omni 1.1 Flash | 4K | 90 | 263 s | 293 s | 8 s |
| Wan 3.0 | 480p | 393 | 158 s | 285 s | 5 s |
| Wan 3.0 | 720p | 2,348 | 195 s | 346 s | 5 s |
| Wan 3.0 | 1080p | 136 | 241 s | 862 s | 6 s |
| Seedance 2.0 | 480p | 140 | 299 s | 446 s | 9 s |
| Seedance 2.0 | 720p | 2,444 | 332 s | 501 s | 5 s |
| Kling 3.0 | 720p | 1,699 | 117 s | 201 s | 5 s |
| Kling 3.0 | 1080p | 516 | 150 s | 294 s | 4 s |
| Gemini Omni 1.2 Flash | 720p | 987 | 67 s | 116 s | 8 s |
| Gemini Omni 1.2 Flash | 1080p | 126 | 100 s | 160 s | 6 s |
| Gemini Omni 1.2 Flash | 4K | 32 | 207 s | 242 s | 8 s |
| Seedance 2.0 Mini | 480p | 122 | 133 s | 173 s | 5 s |
| Seedance 2.0 Mini | 720p | 921 | 163 s | 226 s | 4 s |
| Kling 3.0 Omni | 720p | 403 | 136 s | 176 s | 3 s |
| Kling 3.0 Omni | 1080p | 121 | 141 s | 441 s | 4 s |
| Kling 4.0 | 720p | 376 | 264 s | 514 s | 5 s |
| Kling 4.0 | 1080p | 161 | 293 s | 430 s | 5 s |
| Seedance 2.0 Fast | 480p | 235 | 140 s | 186 s | 5 s |
| Seedance 2.0 Fast | 720p | 223 | 160 s | 221 s | 5 s |
| Wan 3.0 Prime | 720p | 138 | 166 s | 273 s | 8 s |
| Wan 3.0 Prime | 1080p | 119 | 140 s | 265 s | 4 s |
| Veo 3.1 Fast | 720p | 81 | 95 s | 101 s | 6 s |
| Veo 3.1 Fast | 1080p | 118 | 136 s | 157 s | 4 s |
| MiniMax H3 Max | 480p | 42 | 23 s | 26 s | 7 s |
| MiniMax H3 Max | 768p | 126 | 34 s | 62 s | 5 s |
| MiniMax H3 Max | 1080p | 30 | 29 s | 44 s | 15 s |
| Veo 3.1 Lite | 720p | 87 | 78 s | 102 s | 8 s |
| Veo 3.1 Lite | 1080p | 40 | 136 s | 175 s | 6 s |
Success rate per model
A job counts as delivered when the API returned a finished video. Everything else counts as a failure, including content refusals (the model declined the prompt or the picture) and temporary outages. On Revana a failed job costs nothing, but it still costs you time.
Most failures came in bursts, not evenly: Gemini Omni 1.1 Flash delivered 83.4% over the whole period but 92.1% in the last 7 days, and most of its failures were temporary capacity errors from the model's servers. Most of MiniMax H3's failures were jobs that waited too long for capacity in its first weeks; in the last 7 days it delivered all 124 jobs. Content refusals were a small share almost everywhere: for Seedance 2.5 they explain about half of the failures.
When a model fails, the Revana router retries the job on another route that serves the same model, so the numbers above already include that fallback. Read more in AI video API error codes.
Harga per detik
Every video model on Revana is billed per second of output. To compare models with a per-job fee fairly, the chart spreads the fee over a 5 second clip. Prices are list prices for a 5 second clip with sound where the model charges extra for sound.
The cheapest model is not always the best value. A model that is twice the price but delivers on the first try and finishes in half the time can cost less in your product. Compare the price per second with the success rate and the p90 before you choose. For more ways to spend less, see how to cut AI video API costs and the cheapest AI video generation API.
Every video model in one table
Sorted by number of jobs. Each model's page has its full price list, inputs and code; the try link runs it in your browser.
| Model | Res. | Job | Delivered | 7 hari terakhir | p50 | p90 | Median clip | Wait per video second | List price / s | Klip 5 dtk | Try |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Seedance 2.5 | 720p | 48,947 | 99.1% | 99.4% (30,134) | 358 s | 635 s | 9 s | 41.4 s | $0.21 | $1.05 | Try free |
| Gemini Omni 1.1 Flash | 720p | 5,679 | 83.4% | 92.1% (4,392) | 89 s | 160 s | 6 s | 18.2 s | $0.035 + $0.18 a job | $0.355 | Try free |
| Wan 3.0 | 720p | 2,932 | 98.1% | 97.9% (1,887) | 195 s | 346 s | 5 s | 32.5 s | $0.07 | $0.35 | Try free |
| Seedance 2.0 | 720p | 2,673 | 97.2% | 97.4% (2,584) | 332 s | 501 s | 5 s | 58.9 s | $0.15 | $0.75 | Try free |
| Kling 3.0 | 720p + sound | 2,252 | 98.4% | 99.1% (1,488) | 117 s | 201 s | 5 s | 23.7 s | $0.13 | $0.65 | Try free |
| Gemini Omni 1.2 Flash | 720p | 1,250 | 93.0% | 93.2% (1,203) | 67 s | 116 s | 8 s | 9.8 s | $0.072 | $0.36 | Try free |
| Seedance 2.0 Mini | 720p | 1,054 | 99.0% | 99.8% (814) | 163 s | 226 s | 4 s | 31.1 s | $0.05 | $0.25 | Try free |
| Kling 3.0 Omni | 720p + sound | 555 | 94.4% | 94.5% (543) | 136 s | 176 s | 3 s | 34.3 s | $0.13 | $0.65 | Try free |
| Kling 4.0 | 720p | 551 | 97.5% | 97.7% (352) | 264 s | 514 s | 5 s | 46.9 s | $0.135 | $0.675 | Try free |
| Seedance 2.0 Fast | 720p | 477 | 96.0% | 94.0% (316) | 160 s | 221 s | 5 s | 32.2 s | $0.11 | $0.55 | Try free |
| MiniMax H3 | 768p | 356 | 84.8% | 100.0% (124) | 64 s | 1285 s | 5 s | 11.4 s | $0.05 | $0.25 | Try free |
| Wan 3.0 Prime | 720p | 264 | 98.9% | 98.7% (229) | 166 s | 273 s | 8 s | 18.5 s | $0.13 | $0.65 | Try free |
| Veo 3.1 Fast | 720p + sound | 224 | 98.2% | 97.8% (178) | 95 s | 101 s | 6 s | 13.7 s | $0.10 | $0.50 | Try free |
| MiniMax H3 Max | 768p | 200 | 99.0% | 98.6% (144) | 34 s | 62 s | 5 s | 4.8 s | $0.08 | $0.40 | Try free |
| Wan 2.6 I2V Flash | 720p + sound | 166 | 97.0% | 97.0% (165) | 161 s | 447 s | 4 s | 39.8 s | $0.05 | $0.25 | Try free |
| Veo 3.1 | 1080p + sound | 148 | 96.6% | 100.0% (73) | 155 s | 236 s | 6 s | 27.6 s | $0.40 | $2.00 | Try free |
| Veo 3.1 Lite | 720p | 138 | 92.0% | 90.7% (118) | 78 s | 102 s | 8 s | 12.1 s | $0.05 | $0.25 | Try free |
| Kling 2.5 Turbo Pro | std | 135 | 100.0% | 100.0% (135) | 260 s | 767 s | 5 s | 51.7 s | $0.08 | $0.40 | – |
| Hailuo 02 | 512p | 79 | 98.7% | – | 108 s | 226 s | 6 s | 17.8 s | $0.02 | $0.10 | Try free |
| Kling 4.0 Flash | 720p | 55 | 100.0% | 100.0% (55) | 100 s | 181 s | 5 s | 24.2 s | $0.10 | $0.50 | – |
| Grok Imagine Video 1.5 Preview | 480p/720p | 52 | 94.2% | 94.2% (52) | 67 s | 87 s | 4 s | 16.8 s | $0.03 | $0.15 | – |
| Hailuo 02 Standard | 768p | 46 | 100.0% | 100.0% (46) | 262 s | 321 s | 6 s | 44.5 s | $0.04 | $0.20 | – |
| Wan 2.7 | 720p | 41 | 100.0% | 100.0% (39) | 85 s | 150 s | 8 s | 10.8 s | $0.09 | $0.45 | Try free |
| Gemini Omni | 720p | 35 | 85.7% | – | 75 s | 86 s | 4 s | 18.7 s | $0.10 | $0.50 | Try free |
What developers actually generate
Seedance 2.5 carried 72% of all video jobs, even though it is one of the slower and pricier models per second. Developers pick it for quality, long clips (up to 30 seconds), sound, and reference images and video. The second most used model, Gemini Omni 1.1 Flash, is the opposite trade: fast and cheap per second, with a per-job fee.
Bonus: image models
The same measurement for the 14 image models with 30 or more jobs. Nano Banana Pro alone had 74,778 jobs; the fastest was Seedream 5.0 Flash at a median 22 s.
| Model | Ukuran | Job | Delivered | p50 | p90 | List price / image | Try |
|---|---|---|---|---|---|---|---|
| Nano Banana Pro | 2K | 74,778 | 98.9% | 51 s | 135 s | $0.04 | Try free |
| GPT Image 2 | 2K | 36,968 | 99.5% | 48 s | 152 s | $0.03 | Try free |
| GPT Image 2.5 Sunburst | 2K | 18,196 | 99.6% | 35 s | 66 s | $0.02 | Try free |
| Seedream 5.0 Pro | 2K | 4,206 | 99.6% | 108 s | 157 s | $0.04 | Try free |
| Nano Banana 2.1 | 2K | 3,709 | 98.7% | 43 s | 81 s | $0.03 | Try free |
| Nano Banana 2 | 2K | 3,445 | 99.0% | 31 s | 56 s | $0.03 | Try free |
| GPT Image 2.5 Flare | 2K | 2,711 | 98.6% | 46 s | 101 s | $0.02 | Try free |
| GPT Image 2.6 Sunburst | 2K | 1,779 | 99.1% | 50 s | 220 s | $0.02 | Try free |
| GPT Image 2.6 Flare | 2K | 952 | 98.2% | 36 s | 85 s | $0.02 | Try free |
| Nano Banana 2 Lite | 1K | 883 | 97.6% | 37 s | 79 s | $0.02 | Try free |
| Seedream 5.5 | 2K | 529 | 99.8% | 35 s | 61 s | $0.055 | Try free |
| Seedream 5.0 Flash | 2K | 214 | 100.0% | 22 s | 44 s | $0.02 | Try free |
| Seedream 4.5 | 2K | 127 | 94.5% | 35 s | 76 s | $0.04 | Try free |
| Flux 3 Klein | 1K | 110 | 100.0% | 26 s | 45 s | $0.01 | Try free |
How we measured
- Source: every API and dashboard job customers sent to Revana between 20 October and 15 November 2026. Our own team and test accounts are left out.
- Models: every video and image generation model with at least 30 jobs. Talking-avatar, lip-sync, restyle and upscaling tools are left out because they take audio or a source clip and are not comparable. Render times are shown per resolution only where at least 30 jobs finished at that resolution.
- Render time: request accepted to file ready, for delivered jobs. p50 is the median; p90 is the time 9 in 10 jobs finished within.
- Success rate: delivered jobs divided by all finished jobs (delivered plus failed). Cancelled and still-running jobs are not counted.
- Prices: Revana list prices from our model catalogue as of November 2026. Every model has one price whichever route serves it.
- Inputs vary: jobs mix text to video, image to video and reference jobs, and every clip length customers chose. That is the point: this is what real traffic looks like, not one test prompt.
You are free to use these numbers. Please link to this page as the source. We update it every month with the latest 30 days.
To run any of these models yourself, read the API docs or get an API key. One key and one request format work for every model in the table.
Pertanyaan yang sering diajukan
Which AI video API is the fastest?
In our data from 20 October to 15 November 2026, MiniMax H3 Max had the shortest median render time: 34 s at 768p for a median 5 second clip, from the moment the API accepted the request to the moment the file was ready. MiniMax H3 (64 s) and Grok Imagine Video 1.5 Preview (67 s) were next.
Which AI video model is the most reliable?
At volume, Seedance 2.5: 99.1% of 48,947 jobs were delivered (99.5% not counting content refusals). 4 smaller models delivered every job, the largest being Kling 2.5 Turbo Pro (135 of 135).
What is the cheapest AI video API per second?
On Revana in November 2026, Hailuo 02 at 512p costs $0.02 a second, and Grok Imagine Video 1.5 Preview costs $0.03 a second at 720p. Seedance 2.5 is $0.09 at 480p and $0.21 at 720p, with sound.
How long does Seedance 2.5 take to generate a video?
The median Seedance 2.5 job at 720p took 358 s and 9 in 10 finished within 635 s, for a median clip of 9 seconds. At 480p the median was 324 s.
How was render time measured?
From the time the API accepted the request to the time the finished file was stored and ready to download. It includes any queueing and any retry on a second route, so it is what a customer actually waited.
Can I use this data?
Yes. Download the CSV and cite this page as the source. We refresh the same page every month.
Try it first: run Seedance 2.5 free, MiniMax H3 Max free or Kling 3.0 free in your browser. Your first request is free once you save a card at sign-up (saving it charges nothing).
Every model in this benchmark, one API key
Call all 24 video models with the same key and request format, with automatic fallback to another route if one fails. Prepaid credit from $10.