TimeBlind: leaderboard
Metric: Instance accuracy I-Acc (%) over TimeBlind's 600 instances (2,400 video-question trials); an instance counts only when all four trials are answered correctly; TimeBlind minimal pairs: two short videos with the same static content that differ only in temporal structure, each asked two complementary binary or two-choice questions whose answers flip between the videos; videos sampled at 1 FPS, zero-shot, default model configurations; chance is 6.3%; higher is better. Source: arxiv.org. Saturation forecast: Around August 2027. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 48.2 | #77 |
| 2 | GPT-5 | 46.3 | #91 |
| 3 | Gemini 2.5 Pro | 43.5 | #145 |
| 4 | Gemini 2.5 Flash | 33.2 | #237 |
| 5 | Molmo2-8B | 31.2 | #522 |
| 6 | GPT-5 Mini | 30 | #176 |
| 7 | Qwen 3 VL 235B A22B Instruct | 25.8 | #264 |
| 8 | Qwen 3 VL 8B Instruct | 19.7 | #401 |
| 9 | GLM-4.1V-9B (Thinking) | 19.7 | #457 (GLM-4.1V-9B) |
| 10 | Claude Sonnet 4.5 | 13.6 | #138 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=timeblind · How It Works · Data refreshed daily, snapshot 2026-10-11.