BVB: leaderboard
Blender-VideoBench: an agent writes a Blender program that reconstructs a source video, and the render is scored on spatiotemporal fact retention (Dual VQA) and V-JEPA perceptual similarity. Overall is the square-root mean of the two, so a reconstruction has to both look right and be right. 51 model-and-reasoning-effort configurations. Source: arxiv.org/abs/2609.15478.
Metric: Overall (%). Source: arxiv.org. Saturation forecast: Around 2029. 51 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra (High) | 70.07 | #6 (GPT-6 Astra) |
| 2 | GPT-5.6 Sol (xHigh) | 67.49 | #22 (GPT-5.6 Sol) |
| 3 | Grok 4.6 (xHigh) | 67.17 | #29 (Grok 4.6) |
| 4 | Qwen 3.8 Max (High) | 66.24 | #34 (Qwen 3.8 Max) |
| 5 | Claude Opus 5 (High) | 66.21 | #16 (Claude Opus 5) |
| 6 | GPT-5.6 Sol (High) | 65.73 | #22 (GPT-5.6 Sol) |
| 7 | GPT-5.6 Terra (xHigh) | 64.79 | #40 (GPT-5.6 Terra) |
| 8 | GPT-5.6 Terra (High) | 64.76 | #40 (GPT-5.6 Terra) |
| 9 | GPT-5.5 (High) | 64.53 | #33 (GPT-5.5) |
| 10 | GPT-5.6 Sol | 64.09 | #22 |
| 11 | GPT-5.6 Sol (Medium) | 64.01 | #22 (GPT-5.6 Sol) |
| 12 | GPT-5.6 Luna (High) | 63.58 | #79 (GPT-5.6 Luna) |
| 13 | Muse Spark 1.3 (xHigh) | 63.4 | #31 (Muse Spark 1.3) |
| 14 | GPT-5.6 Sol (Low) | 63.23 | #22 (GPT-5.6 Sol) |
| 15 | Grok 4.5 (High) | 62.81 | #54 (Grok 4.5) |
No result here: #3 Claude Opus 5.5, #7 GPT-6.1 Sol, #9 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=bvb · How It Works · Data refreshed daily, snapshot 2026-11-05.