ExtremeWhenBench: leaderboard
ExtremeWhenBench (2026) tests natural-language temporal grounding in hour-scale videos: for each of 2,273 open-form queries over 194 videos averaging 76 minutes, a model returns the start and end time of the described moment. The board is mean temporal IoU scaled to 0-100 on full-length videos, each model at its best frame budget within its context limit and parse failures scored 0; higher is better.
Metric: Mean IoU x 100 (0-100) between the predicted and gold time intervals over the 2,273 queries on full hour-scale videos, each model at its best frame budget within its context limit (GPT-5.4: 64 frames as 8x8 grids; Gemini: raw video), parse failures scored 0; higher is better. Source: arxiv.org. Saturation forecast: Around 2031. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.5 Flash | 11.5 | #101 |
| 2 | Qwen 3.5 9B | 11 | #803 |
| 3 | Gemini 2.5 Flash | 5.3 | #453 |
| 4 | GPT-5.4 | 1.3 | #123 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=extremewhenbench · How It Works · Data refreshed daily, snapshot 2026-11-13.