Benchmark Almanac

4DCodeBench

GPT-6 Astra 0.79 leads since October 2026
Not in model cards

Agents reconstruct dynamic scenes from video as executable graphics programs, including physical simulations, so rendered output matches the observed motion.

200 GitHub stars at November 1st, 2026.

Released
October 2026
Kind
Physical prediction
Level
Expert
Scored
Checked automatically
Answers
Code
Tasks
All public
Strongest model tested
GPT-6 Astra #3 on the capability index
Last model added
September 2026
Scores refreshed
November 12th, 2026

In the authors’ words

We introduce 4DCodeBench, a benchmark for 4D inverse graphics through code generation, in which agents reconstruct dynamic scenes from video as executable graphics programs. To accomplish this, agents must translate visual observations into compact representations of scene structure and dynamics, by implementing abstractions such as physical simulations to reproduce complex behavior. To evaluate this capability, we curate a set of real-world videos and construct synthetic scenes spanning diverse physical phenomena, including deformation, fluid flow, and fracture. We perform extensive benchmarking of frontier models, finding that strong static reconstruction capabilities do not yet translate into reliable reconstruction of complex dynamics. 4DCodeBench provides a testbed for tracking progress toward agents that can interpret the dynamics of the world through code. Our benchmark is available at https://github.com/4DCodeBench/4DCodeBench

Abstract of 4DCodeBench: Benchmarking Agents on Inverse Graphics of Dynamic Scenes (arXiv:2610.03715, CC0)
16
Models scored on its own leaderboard

GPT-6 Astra leads with 0.79 among models released by September 2026

Best score per model from the benchmark's own leaderboard, read on November 12th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.

The newest model with a score here came out in September 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

16 models. Scroll sideways for the rest.

  1. 0.79 GPT-6 Astra
  2. 0.78 Claude Opus 5.5
  3. 0.74 Claude Fable 5.1
  4. 0.72 Claude Opus 5
  5. 0.67 GPT-5.6 Luna
  6. 0.65 GPT-5.6 Sol
  7. 0.64 Gemini 3.8 Flash
  8. 0.60 DeepSeek V4.1 Flash
  9. 0.56 Qwen3.8 Flash
  10. 0.55 GPT-5.6 Terra
  11. 0.47 GLM-5.3-Flash
  12. 0.44 Muse Glimmer
  13. 0.43 MiMo-V2.5
  14. 0.34 MiniMax-M3
  15. 0.32 Mistral Medium 3.5
  16. 0.19 Gemma 4 31B IT
  1. 1 GPT-6 Astra Max 0.79
  2. 2 Claude Opus 5.5 High 0.78
  3. 3 Claude Fable 5.1 High 0.74
  4. 4 Claude Opus 5 High 0.72
  5. 5 GPT-5.6 Luna Max 0.67
  6. 6 GPT-5.6 Sol High 0.65
  7. 7 Gemini 3.8 Flash High 0.64
  8. 8 DeepSeek V4.1 Flash High 0.60
  9. 9 Qwen3.8 Flash XHigh 0.56
  10. 10 GPT-5.6 Terra High 0.55
  11. 11 GLM-5.3-Flash Max 0.47
  12. 12 Muse Glimmer High 0.44
  13. 13 MiMo-V2.5 Default 0.43
  14. 14 MiniMax-M3 Default 0.34
  15. 15 Mistral Medium 3.5 0.32
  16. 16 Gemma 4 31B IT 0.19

Source: Official 4DCodeBench leaderboard

#ModelScoreSettingDateSource
1 GPT-6 Astra 0.79 Max, Codex Sep 2026 Official leaderboard
2 Claude Opus 5.5 0.78 High, Claude Code Sep 2026 Official leaderboard
3 Claude Fable 5.1 0.74 High, Claude Code Sep 2026 Official leaderboard
4 Claude Opus 5 0.72 High, Claude Code Jul 2026 Official leaderboard
5 GPT-5.6 Luna 0.67 Max, Codex Jul 2026 Official leaderboard
6 GPT-5.6 Sol 0.65 High, Codex Jul 2026 Official leaderboard
7 Gemini 3.8 Flash 0.64 High, Antigravity Sep 2026 Official leaderboard
8 DeepSeek V4.1 Flash 0.60 High, Stirrup Sep 2026 Official leaderboard
9 Qwen3.8 Flash 0.56 XHigh, Stirrup Aug 2026 Official leaderboard
10 GPT-5.6 Terra 0.55 High, Codex Jul 2026 Official leaderboard
11 GLM-5.3-Flash 0.47 Max, Stirrup Aug 2026 Official leaderboard
12 Muse Glimmer 0.44 High, Stirrup Aug 2026 Official leaderboard
13 MiMo-V2.5 0.43 Default, Stirrup Apr 2026 Official leaderboard
14 MiniMax-M3 0.34 Default, Stirrup Jun 2026 Official leaderboard
15 Mistral Medium 3.5 0.32 Stirrup Apr 2026 Official leaderboard
16 Gemma 4 31B IT 0.19 Stirrup Apr 2026 Official leaderboard

4DCodeBench score versus cost

The line joins the runs that no other run beats on both measures. We redrew this chart from the numbers on the official leaderboard.

0 0.2 0.4 0.6 0.8 $0.01$0.1$1$10$100 Mean cost per task ($), log scale, lower is better Overall score (score) Mistral Medium 3.5: Overall score 0.32 score, mean cost per task $3.57 MiniMax-M3 (Default): Overall score 0.34 score, mean cost per task $2.47 GPT-5.6 Sol (High): Overall score 0.65 score, mean cost per task $2.47 GPT-5.6 Terra (High): Overall score 0.55 score, mean cost per task $0.62 Claude Opus 5 (High): Overall score 0.72 score, mean cost per task $17.5 GLM-5.3-Flash (Max): Overall score 0.47 score, mean cost per task $6.28 Qwen3.8 Flash (XHigh): Overall score 0.56 score, mean cost per task $2.77 Claude Fable 5.1 (High): Overall score 0.74 score, mean cost per task $15.1 Gemini 3.8 Flash (High): Overall score 0.64 score, mean cost per task $2.72 GPT-6 Astra (High): Overall score 0.77 score, mean cost per task $6.28 Gemma 4 31B IT: Overall score 0.19 score, mean cost per task $0.08 MiMo-V2.5 (Default): Overall score 0.43 score, mean cost per task $0.16 GPT-5.6 Luna (Max): Overall score 0.67 score, mean cost per task $0.67 Muse Glimmer (High): Overall score 0.44 score, mean cost per task $0.29 GPT-6 Astra (Max): Overall score 0.79 score, mean cost per task $11.3 GPT-6 Astra (Low): Overall score 0.73 score, mean cost per task $3.5 DeepSeek V4.1 Flash (High): Overall score 0.6 score, mean cost per task $0.44 Claude Opus 5.5 (High): Overall score 0.78 score, mean cost per task $6.27 GPT-6 Astra (Max)Claude Opus 5.5 (High)GPT-6 Astra (Low)GPT-5.6 Luna (Max)DeepSeek V4.1 Flash (High)Muse Glimmer (High)MiMo-V2.5 (Default)Gemma 4 31B IT 0 0.2 0.4 0.6 0.8 $0.01$0.1$1$10$100 Mean cost per task ($), log scale, lower is better Overall score (score) Mistral Medium 3.5: Overall score 0.32 score, mean cost per task $3.57 MiniMax-M3 (Default): Overall score 0.34 score, mean cost per task $2.47 GPT-5.6 Sol (High): Overall score 0.65 score, mean cost per task $2.47 GPT-5.6 Terra (High): Overall score 0.55 score, mean cost per task $0.62 Claude Opus 5 (High): Overall score 0.72 score, mean cost per task $17.5 GLM-5.3-Flash (Max): Overall score 0.47 score, mean cost per task $6.28 Qwen3.8 Flash (XHigh): Overall score 0.56 score, mean cost per task $2.77 Claude Fable 5.1 (High): Overall score 0.74 score, mean cost per task $15.1 Gemini 3.8 Flash (High): Overall score 0.64 score, mean cost per task $2.72 GPT-6 Astra (High): Overall score 0.77 score, mean cost per task $6.28 Gemma 4 31B IT: Overall score 0.19 score, mean cost per task $0.08 MiMo-V2.5 (Default): Overall score 0.43 score, mean cost per task $0.16 GPT-5.6 Luna (Max): Overall score 0.67 score, mean cost per task $0.67 Muse Glimmer (High): Overall score 0.44 score, mean cost per task $0.29 GPT-6 Astra (Max): Overall score 0.79 score, mean cost per task $11.3 GPT-6 Astra (Low): Overall score 0.73 score, mean cost per task $3.5 DeepSeek V4.1 Flash (High): Overall score 0.6 score, mean cost per task $0.44 Claude Opus 5.5 (High): Overall score 0.78 score, mean cost per task $6.27 GPT-6 Astra (Max)Claude Opus 5.5 (High)GPT-6 Astra (Low)GPT-5.6 Luna (Max)DeepSeek V4.1 Flash (High)Muse Glimmer (High)MiMo-V2.5 (Default)Gemma 4 31B IT
ModelSettingOverall scoreMean cost per task
GPT-6 Astra Max, Codex 0.79 score $11.32
Claude Opus 5.5 High, Claude Code 0.78 score $6.27
GPT-6 Astra High, Codex 0.77 score $6.28
Claude Fable 5.1 High, Claude Code 0.74 score $15.07
GPT-6 Astra Low, Codex 0.73 score $3.5
Claude Opus 5 High, Claude Code 0.72 score $17.52
GPT-5.6 Luna Max, Codex 0.67 score $0.67
GPT-5.6 Sol High, Codex 0.65 score $2.47
Gemini 3.8 Flash High, Antigravity 0.64 score $2.72
DeepSeek V4.1 Flash High, Stirrup 0.6 score $0.44
Qwen3.8 Flash XHigh, Stirrup 0.56 score $2.77
GPT-5.6 Terra High, Codex 0.55 score $0.62
GLM-5.3-Flash Max, Stirrup 0.47 score $6.28
Muse Glimmer High, Stirrup 0.44 score $0.29
MiMo-V2.5 Default, Stirrup 0.43 score $0.16
MiniMax-M3 Default, Stirrup 0.34 score $2.47
Mistral Medium 3.5 Stirrup 0.32 score $3.57
Gemma 4 31B IT Stirrup 0.19 score $0.08

The best score, 0.79, costs $11 a task

Runs on the cost frontier score higher than any run that's cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.

Google DeepMind Xiaomi Corp OpenAI Meta DeepSeek Anthropic Cost frontier One model at each effort level
0.00 0.20 0.40 0.60 0.80 $0.01 $0.10 $1.00 $10 $100 Cost per task, log scale GPT-6 Astra, Max GPT-6 Astra, Low GPT-5.6 Luna, Max DeepSeek V4.1 Flash, High Muse Glimmer, High Gemma 4 31B IT 0.00 0.20 0.40 0.60 0.80 $0.01 $0.10 $1.00 $10 $100 Cost per task, log scale GPT-6 Astra, Max GPT-6 Astra, Low

Source: Official 4DCodeBench leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
GPT-6 Astra Max Codex 0.79 $11 67,475
Claude Opus 5.5 High Claude Code 0.78 $6.27 113,408
GPT-6 Astra High Codex 0.77 $6.28 29,279
Claude Fable 5.1 High Claude Code 0.74 $15 140,433
GPT-6 Astra Low Codex 0.73 $3.50 13,439
Claude Opus 5 High Claude Code 0.72 $18 253,626
GPT-5.6 Luna Max Codex 0.67 $0.67 106,404
GPT-5.6 Sol High Codex 0.65 $2.47 26,032
Gemini 3.8 Flash High Antigravity 0.64 $2.72 206,201
DeepSeek V4.1 Flash High Stirrup 0.60 $0.44 134,310
Qwen3.8 Flash XHigh Stirrup 0.56 $2.77 1,505,896
GPT-5.6 Terra High Codex 0.55 $0.62 17,283
GLM-5.3-Flash Max Stirrup 0.47 $6.28 451,636
Muse Glimmer High Stirrup 0.44 $0.29 45,242
MiMo-V2.5 Default Stirrup 0.43 $0.16 153,926
MiniMax-M3 Default Stirrup 0.34 $2.47 144,665
Mistral Medium 3.5 Stirrup 0.32 $3.57 36,589
Gemma 4 31B IT Stirrup 0.19 $0.08 19,330

Number one over time

Lab claims count from the day the card came out, independent runs from the model’s release.

OpenAI
Own leaderboard 1 leader
GPT-6 Astra 0.79

Source: Official 4DCodeBench leaderboard

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

On the benchmark’s own leaderboard
0.00 0.20 0.40 0.60 0.80 2026 0.00 0.20 0.40 0.60 0.80 2026

Source: Official 4DCodeBench leaderboard

No lab model card reports it yet

We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark.

More in World models

All 39 in World models
  • AgentWorldBench In use Language world models predict the next environment observation after an agent action, in 7 domains from MCP and search to Android and desktop OS. A judge model scores each prediction against the real observation on five dimensions. GPT-5.4 58.3
  • IntPhys 2 In use Models watch simulated videos and judge whether each event is physically possible, testing object permanence, immutability, continuity and solidity. Scored by accuracy. Muse Spark Open 68.3%
  • QwenWorldBench Fading Models act as world models that simulate agent environments in 7 domains, such as terminal, web, and Android. A rubric judge scores five dimensions. Qwen3.7-Plus 62.1%
  • VPCT Fading Models view an image of a ball, ramps, and three buckets and predict which bucket the ball falls into. Scored by accuracy on 100 problems. Gemini 3 Pro 91%
  • ContPhy Fading Models infer physical properties such as mass and density of fluids and soft bodies from videos and predict their dynamics. Seed2.0 Pro 67.4%
  • RBench Only in system cards 650 image+text prompts for robot video generation (tasks and embodiments). MLLM judge scores task completion, physical plausibility, stability. Wan 2.6 60.7%