MEDit-Bench - F1@0.5: leaderboard
MEDit-Bench tests message-driven video editing: given a long video and an editing message, a model must choose the sequence of cuts that tells that story, and its edit is compared with the cuts professional editors made for the same message. This board is cut-level F1 after matching predicted and reference cuts one to one at a temporal intersection over union of at least 0.5, so it also rewards the right number and placement of cuts.
Metric: F1@0.5 (%; cut-level F1 after one-to-one matching of predicted and reference cuts at temporal IoU of at least 0.5, averaged over the 540 video, message and reference-edit triplets (60 long-form videos, three editing messages each, three professional edits per message); zero-shot: the model maps the video and the message to a sequence of cuts). Source: arxiv.org. Saturation forecast: Around 2030. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 15.2 | #137 |
| 2 | GPT-5 | 15.1 | #167 |
| 3 | GPT-5 Mini | 11.2 | #348 |
| 4 | Gemini 2.5 Pro | 10.2 | #266 |
| 5 | Gemma 4 E4B (Non-reasoning) | 7.1 | #2214 (Gemma 4 E4B) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=medit-bench-f1-0-5 · How It Works · Data refreshed daily, snapshot 2026-11-13.