4DCodeBench: leaderboard

Agentic reconstruction of 200 videos (100 real and 100 simulated) as executable graphics programs. The score is the equal-weight mean of five reconstruction families on fixed [0,1] scales: perception, 2D motion, depth, 3D shape and 3D motion. Failed renders incur declared penalties. Models use the published provider-specific coding harnesses and effort settings; VLM preference Elo and VQA are separate measurements.

Metric: Reconstruction score (0–1), five-family mean over 200 scenes; provider-specific coding harnesses. Source: 4dcodebench.com. Saturation forecast: Around February 2027. 16 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-6 AstraMax0.7908#11
2Claude Opus 5.5High0.7784#9
3Claude Fable 5.1High0.741#20
4Claude Opus 5High0.7172#26
5GPT-5.6 LunaMax0.6695#135
6GPT-5.6 SolHigh0.6454#36
7Gemini 3.8 FlashHigh0.6394#43
8DeepSeek V4.1 FlashHigh0.5965#69
9Qwen 3.8 FlashxHigh0.5627#132
10GPT-5.6 TerraHigh0.5505#74
11GLM-5.3 FlashMax0.4732#137
12Muse Glimmer 30BHigh0.4419#395
13MiMo-V2.50.4348#387
14MiniMax-M30.3394#290
15Mistral Medium 3.50.3152#742

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=4dcodebench · How It Works · Data refreshed daily, snapshot 2026-11-22.