SceneActBench: leaderboard
Benchmark for visually conditioned action on complete multi-object 3D scenes: VLM agents drive headless Blender through a shared MCP tool interface across five tasks (layout, camera pose, articulation, reconstruction, dynamics) built from 210 source instances yielding 520 task cases, with each final output scored against hidden 3D ground truth by geometric metrics rather than by text. Overall aggregates the five tasks.
Metric: Overall Score (%). Source: github.com. Saturation forecast: Around 2029. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Seed 2.0 Pro (High) | 50.2 | #122 (Seed 2.0 Pro) |
| 2 | Claude Opus 4.6 (High) | 48.9 | #67 (Claude Opus 4.6) |
| 3 | GPT-5.4 (High) | 48.7 | #83 (GPT-5.4) |
| 4 | GPT-5.4 (Medium) | 48.7 | #83 (GPT-5.4) |
| 5 | Gemini 3.1 Pro (Preview) (High) | 45.4 | #64 (Gemini 3.1 Pro (Preview)) |
| 6 | Kimi K2.6 (Thinking) | 41.2 | #113 (Kimi K2.6) |
| 7 | Step 3.7 Flash (High) | 41.1 | #247 (Step 3.7 Flash) |
| 8 | Claude Sonnet 5 (High) | 39.5 | #65 (Claude Sonnet 5) |
| 9 | MiniMax-M3 (High) | 38.6 | #151 (MiniMax-M3) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=sceneactbench · How It Works · Data refreshed daily, snapshot 2026-11-05.