Benchmarks
351 sources · 1,261 models × 3,053 benchmarks · 160,189 scoresVibePrevisBench gives coding agents 242 tasks in which they build an animated, usually multi-shot 3D scene in Blender from a script, photos, a continuation or a reference video, and render a proxy video. The score averages a deterministic check of the exported scene against hidden ground-truth facts and a Gemini 4 judge's check that those facts read on screen; a failed run scores 0.
163rd strongest of 803 fitted benchmarks
Signal-to-noise +11.05
434th hardest of 803 fitted benchmarks
Capability needed ≲ 147.9
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | GPT-6 Astra | max · Codex | 77.7 | |
| 2 | GPT-6.1 Sol | max · Codex | 76.1 | |
| 3 | Gemini 4 Argon | high · Claude Code | 75.0 | |
| 4 | Claude Opus 5.5 | high · Claude Code | 74.1 | |
| 5 | GPT-5.6 Sol | max · Codex | 70.2 | |
| 6 | DeepSeek V4.1 Flash 1017 | max · Claude Code | 69.6 | |
| 7 | Gemini 3.8 Flash | high · Claude Code | 64.7 | |
| 8 | GLM-5.3 Flash | high · Claude Code | 63.7 | |
| 9 | Muse Spark 1.3 | x-high · Claude Code | 61.1 |