MTAVG-Bench: leaderboard
Metric: Average score (%) over the nine failure dimensions, the unweighted mean of the dimension scores; MTAVG-Bench: 2,410 human-verified diagnostic questions on 1,880 multi-speaker dialogue videos generated by Veo 3.1, Sora 2 and Wan 2.5, each question grounded in an annotated generation failure; single-answer and pairwise-comparison questions score 1 when correct, multiple-answer questions the share of gold options selected; the omni-modal model receives the video with its audio and the generation prompt; higher is better. Source: arxiv.org. Saturation forecast: Around March 2028. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 56.84 | #77 |
| 2 | Gemini 2.5 Pro | 54.19 | #145 |
| 3 | Gemini 3 Flash | 53.02 | #93 |
| 4 | Gemini 2.5 Flash (Thinking) | 50.38 | #237 (Gemini 2.5 Flash) |
| 5 | Gemini 2.5 Flash (Non-reasoning) | 49.75 | #237 (Gemini 2.5 Flash) |
| 6 | Qwen2.5-Omni-7B | 42.81 | #601 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=mtavg-bench · How It Works · Data refreshed daily, snapshot 2026-10-11.