VideoVIBE: leaderboard
VideoVIBE (2026) turns human-operated recordings of one-shot generated interactive websites into multiple-choice questions about which failure a session shows, answered by video MLLMs from the recording and the page source. Unweighted mean accuracy over twelve failure subdimensions in percent; higher is better.
Metric: Failure-diagnosis accuracy (%; unweighted macro-average of the twelve subdimension accuracies; zero-shot single-choice diagnosis from the interaction video and the page source code (Video+HTML), about 1.7K questions built from 6,338 verified failures of one-shot generated web applications recorded under human operation; free-form answers mapped deterministically to an option). Source: arxiv.org. Saturation forecast: Around August 2027. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 64.54 | #453 |
| 2 | Gemini 3.1 Pro (Preview) | 64.25 | #94 |
| 3 | Qwen 3.5 35B A3B | 63.9 | #482 |
| 4 | Qwen 3 VL 30B A3B (Thinking) | 60.96 | #712 |
| 5 | Qwen 3 VL 8B (Thinking) | 59.92 | #887 |
| 6 | Gemini 3.1 Flash Lite | 58.21 | #364 |
| 7 | Gemini 3 Flash | 57.09 | #164 |
| 8 | GLM-4.6V | 53.78 | #701 |
| 9 | Qwen 3 VL 30B A3B Instruct | 52.27 | #871 |
| 10 | MiMo-V2.5 | 49.61 | #354 |
| 11 | Qwen 3 VL 8B Instruct | 45.69 | #981 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=videovibe · How It Works · Data refreshed daily, snapshot 2026-11-13.