SceneActBench - Dynamic: leaderboard

Metric: Task score (0-100): 10 low-poly Kenney scenes from a sampled 144-frame video: place and keyframe every mover, scored by mean motion and layout error against unit references; every configuration drives one headless Blender through the same MCP tool interface (scene and object inspection, Python execution, rendering) with task step budgets; each native geometric error or F-score is mapped to a 0 to 100 case score against a fixed reference, invalid outputs score 0, cases are averaged per task. Source: arxiv.org. Saturation forecast: Around August 2028. 11 models tracked.

Top models

#ModelScore
1Seed 2.0 Pro (High)70.7
2GPT-5.4 (Medium)68.5
3Gemini 3.1 Pro (Preview) (High)63.9
4Claude Opus 4.6 (High)63.2
5Step 3.7 Flash (High)57.9
6Claude Sonnet 5 (High)49.9
7GPT-5.4 (High)46.7
8Kimi K2.6 (Thinking)44.8
9MiniMax M3 (High)41.4

Interactive version: theaggregate.ai/benchmark?slug=sceneactbench-dynamic · How It Works · Data refreshed daily, snapshot 2026-09-29.