Research
What actually happens when people generate AI video: 1652 real generations, 11 models
Between 2026-01-29 and 2026-11-09, 1110 people made 1652 text- and image-to-video generations on Viralance. This page reports, per model, how often a generation finished, why the rest did not, and how long people waited. It is observational data from real use, not a controlled quality test. Reuse it with a link to this page (CC BY 4.0).
Key findings
- real generations
- 1652
- of generations finished
- 89%
- failures were content-policy refusals
- 138 of 176
- fastest median wait (MiniMax H3 Max Turbo)
- 1.3 min
- 89% of generations finished. Of the 176 that did not, 138 (78%) were refused by the model provider's content policy; 14 timed out and 24 failed for other technical reasons.
- Most reliable (lower bound of the 95% interval ≥ 80%): Hailuo 2.3 100%, Kling 3.0 Pro 100%, VEO 3.1 Fast 98%, VEO 3.1 Lite 92%.
- Content-policy refusals vary a lot by model, from 1% on VEO 3.1 Fast to 19% on Grok Imagine, 22% on MiniMax H3, 24% on Seedance 2.0, 56% on Grok Imagine 1.5. These rates depend on what people asked for, so read them as "how often real prompts got refused here", not as a measure of model strictness.
- Typical wait is 2–5 minutes. Fastest median: MiniMax H3 Max Turbo (1.3 min); slowest median: Seedance 2.0 (4.5 min). One in ten generations took longer than the p90 column below.
- VEO 3.1 Lite was charged 0.7 credits per second of output vs 1.46 for VEO 3.1 Fast (52% less), with a 92% vs 98% completion rate.
Per-model results
Models with at least 15 generations. Success = the video was delivered. The 95% interval shows how uncertain each rate is with this sample size. Credits per second are what was charged at the time of each generation (prices changed during the window; see current prices).
| Model | Generations | Success | 95% interval | Policy refusals | Median wait | p90 wait | Credits / output sec | Image-to-video |
|---|---|---|---|---|---|---|---|---|
| VEO 3.1 Lite | 1002 | 92% | 90%–93% | 66 | 2.0 min | 3.1 min | 0.7 | 50% |
| VEO 3.1 Fast | 149 | 98% | 94%–99% | 2 | 2.3 min | 3.0 min | 1.46 | 28% |
| MiniMax H3 Max Turbo | 121 | 83% | 76%–89% | 18 | 1.3 min | 1.7 min | 0.49 | 57% |
| Grok Imagine | 79 | 77% | 67%–85% | 15 | 2.3 min | 3.6 min | 0.61 | 89% |
| MiniMax H3 | 67 | 76% | 65%–85% | 15 | 2.7 min | 4.2 min | 0.67 | 52% |
| PixVerse V6 | 36 | 92% | 78%–97% | 0 | 2.1 min | 2.5 min | 0.89 | 83% |
| Hailuo 2.3 | 34 | 100% | 90%–100% | 0 | 3.4 min | 7.5 min | 0.52 | 68% |
| Seedance 2.0 | 29 | 76% | 58%–88% | 7 | 4.5 min | 9.6 min | 2.96 | 48% |
| Kling 3.0 Pro | 28 | 100% | 88%–100% | 0 | 4.5 min | 7.4 min | 1.96 | 25% |
| Kling 3.0 Standard | 26 | 92% | 76%–98% | 2 | 2.7 min | 4.4 min | 1.59 | 73% |
| Grok Imagine 1.5 | 16 | 44% | 23%–67% | 9 | 2.1 min | 2.1 min | 1.07 | 100% |
Method and limits
Source: Viralance's generation records. Included: text-to-video and image-to-video jobs from regular accounts. Excluded: test, admin and QA accounts; Viralance's own internal accounts, including the account we use for API and MCP test runs; accounts flagged for abuse or banned; video-to-video tools (upscale, reframe, motion transfer); and 11 failures caused by Viralance integration bugs (a missing endpoint mapping in July and a request-parsing error in April–May), because those measure us, not the model. Only aggregate numbers are published: no prompts, media or account data.
Correction, 1 October 2026: the first version of this page (data to 25 September) counted our internal API and MCP test account as a regular user. Those runs are now excluded and every figure on this page and in the CSV was recalculated; models that no longer have 15 generations were removed from the table.
Wait time runs from job submission to the moment Viralance marked the video ready: the provider's queue and generation, plus a short delivery hold (about 45 seconds) and up to a minute of status-check delay, because the completion time is recorded by a once-a-minute server check. Failure reasons come from provider error messages: "policy" covers content/safety refusals, "timeout" covers jobs that exceeded the time limit. Failed generations on Viralance are refunded automatically.
Limits: users chose their own models, prompts, durations and resolutions, so models were not tested on the same inputs; popular models have tighter intervals than rare ones; this says nothing about visual quality. For a same-prompt comparison with measured output files, see our 19-model delivery benchmark.
Cite as: Viralance, "Real-World AI Video Generation Data 2026", 2026-11-09, https://viralance.com/research/real-world-ai-video-generation-2026