Benchmarks
351 sources · 1,261 models × 3,053 benchmarks · 160,189 scoresVibePrevisBench's judged half: a Gemini 4 judge checks whether each task's ground-truth facts read on screen in the rendered proxy video, averaged over the 242 tasks. The paper finds the judge favours Gemini models slightly and other judges score lower, so it is shown as a diagnostic.
No signal or difficulty estimate. Held out of the fit by design.
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | GPT-6 Astra | max · Codex | 85.4 | |
| 2 | GPT-6.1 Sol | max · Codex | 83.1 | |
| 3 | Gemini 4 Argon | high · Claude Code | 81.7 | |
| 4 | Claude Opus 5.5 | high · Claude Code | 79.8 | |
| 5 | DeepSeek V4.1 Flash 1017 | max · Claude Code | 77.9 | |
| 6 | GPT-5.6 Sol | max · Codex | 77.5 | |
| 7 | GLM-5.3 Flash | high · Claude Code | 69.5 | |
| 8 | Gemini 3.8 Flash | high · Claude Code | 68.7 | |
| 9 | Muse Spark 1.3 | x-high · Claude Code | 65.0 |