3DSceneBench
Agents reconstruct interactive MuJoCo scenes from images and task descriptions. Verifiers measure appearance, geometry, and physical functionality, then average the three scores.
The full suite contains 49 scenes. The public mini package contains seven scenes. Full-suite leaderboard scores must not be presented as mini scores.
- Made by
- Prana Labs and University of California, Berkeley
- Released
- November 2026
- Kind
- Physical prediction
- Level
- Expert
- Scored
- Checked automatically
- Answers
- Agent task
- Tasks
- Public, plus a private test set
- Measures
- Capability
- Language
- English
- Interaction
- Agentic
- Environment
- MuJoCo
- Items from
- Generated
- Refresh
- One fixed set
- Run by
- A host
- In views
- Agents, Vision, Image, video and 3D generation, Robotics
- Strongest model tested
- GPT-6 Astra #4 on the capability index
- Last model added
- October 2026
- Scores refreshed
- November 20th, 2026
GPT-6 Astra leads with 59.5
Best score per model from the benchmark's own leaderboard, read on November 20th, 2026. Where a model ran with several settings or harnesses, we keep its best.
- 1 GPT-6 Astra L0; max effort 59.5
- 2
Gemini 4 Argon L0; high effort 56.9
- 3 Claude Opus 5.5 L0; max effort 56.8
- 4 Kimi K3.1 Pro L0; high effort 37.1
- 5
DeepSeek V4.1 Flash L0; high effort 30.6
- 6
MiniMax-M3.1-Flash L0; high effort 28.8
- 7
Qwen 3.8 Max L0; high effort 26.1
- 8 MiMo-V2.6 Pro L0; high effort 25.0
- 9 GLM-5.3-Flash L0; high effort 24.5
- 10
Mistral Large 4 L0; high effort 24.0
Source: Official 3DSceneBench leaderboard
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-6 Astra | 59.5 | L0; max effort | Sep 2026 | Official leaderboard |
| 2 | | 56.9 | L0; high effort | Sep 2026 | Official leaderboard |
| 3 | Claude Opus 5.5 | 56.8 | L0; max effort | Sep 2026 | Official leaderboard |
| 4 | Kimi K3.1 Pro | 37.1 | L0; high effort | Oct 2026 | Official leaderboard |
| 5 | | 30.6 | L0; high effort | Sep 2026 | Official leaderboard |
| 6 | | 28.8 | L0; high effort | Oct 2026 | Official leaderboard |
| 7 | | 26.1 | L0; high effort | Aug 2026 | Official leaderboard |
| 8 | MiMo-V2.6 Pro | 25.0 | L0; high effort | Sep 2026 | Official leaderboard |
| 9 | GLM-5.3-Flash | 24.5 | L0; high effort | Aug 2026 | Official leaderboard |
| 10 | | 24.0 | L0; high effort | Oct 2026 | Official leaderboard |
Rankings by part
We read 12 parts from the benchmark’s own leaderboard, each ranking models on one part of the test, such as Appearance, Geometry and Functionality. Gemini 4 Argon leads 7 of them, and 2 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
Appearance
10 models we track have a score here.
- 1 GPT-6 Astra 79.3
- 2 Claude Opus 5.5 ↑1 70.5
- 3
Gemini 4 Argon ↓1 66.7
- 4 Kimi K3.1 Pro 55.3
- 5
MiniMax-M3.1-Flash ↑1 48.2
- 6
DeepSeek V4.1 Flash ↓1 45.1
- 7 MiMo-V2.6 Pro ↑1 40.0
- 8
Qwen 3.8 Max ↓1 39.0
- 9 GLM-5.3-Flash 36.9
- 10
Mistral Large 4 36.2
Geometry
10 models we track have a score here.
- 1
Gemini 4 Argon ↑1 62.1
- 2 Claude Opus 5.5 ↑1 58.1
- 3 GPT-6 Astra ↓2 58.0
- 4 Kimi K3.1 Pro 35.0
- 5
DeepSeek V4.1 Flash 29.1
- 6
Qwen 3.8 Max ↑1 26.1
- 7
MiniMax-M3.1-Flash ↓1 25.2
- 8
Mistral Large 4 ↑2 24.7
- 9 GLM-5.3-Flash 23.9
- 10 MiMo-V2.6 Pro ↓2 23.5
Functionality
10 models we track have a score here.
- 1 Claude Opus 5.5 ↑2 42.1
- 2
Gemini 4 Argon 42.0
- 3 GPT-6 Astra ↓2 41.1
- 4 Kimi K3.1 Pro 20.9
- 5
DeepSeek V4.1 Flash 17.7
- 6
Qwen 3.8 Max ↑1 13.3
- 7
MiniMax-M3.1-Flash ↓1 12.9
- 8 GLM-5.3-Flash ↑1 12.7
- 9 MiMo-V2.6 Pro ↓1 11.6
- 10
Mistral Large 4 11.1
Inventory
10 models we track have a score here.
- 1
Gemini 4 Argon ↑1 76.0
- 2 Claude Opus 5.5 ↑1 70.9
- 3 GPT-6 Astra ↓2 69.5
- 4 Kimi K3.1 Pro 47.0
- 5
DeepSeek V4.1 Flash 43.9
- 6
Mistral Large 4 ↑4 41.2
- 7
MiniMax-M3.1-Flash ↓1 38.0
- 8
Qwen 3.8 Max ↓1 36.5
- 9 GLM-5.3-Flash 36.0
- 10 MiMo-V2.6 Pro ↓2 34.9
Shape
10 models we track have a score here.
- 1 GPT-6 Astra 69.3
- 2 Claude Opus 5.5 ↑1 68.4
- 3
Gemini 4 Argon ↓1 68.2
- 4 Kimi K3.1 Pro 59.7
- 5
Qwen 3.8 Max ↑2 53.0
- 6
MiniMax-M3.1-Flash 52.7
- 7
DeepSeek V4.1 Flash ↓2 52.4
- 8 MiMo-V2.6 Pro 51.3
- 9 GLM-5.3-Flash 46.5
- 10
Mistral Large 4 43.7
Pose and size
10 models we track have a score here.
- 1
Gemini 4 Argon ↑1 48.4
- 2 GPT-6 Astra ↓1 41.1
- 3 Claude Opus 5.5 39.8
- 4 Kimi K3.1 Pro 8.50
- 5
Mistral Large 4 ↑5 5.10
- 6
MiniMax-M3.1-Flash 4.50
- 7
DeepSeek V4.1 Flash ↓2 4.40
- 8 GLM-5.3-Flash ↑1 4.30
- 9
Qwen 3.8 Max ↓2 3.20
- 10 MiMo-V2.6 Pro ↓2 2.40
Relations
10 models we track have a score here.
- 1
Gemini 4 Argon ↑1 56.6
- 2 GPT-6 Astra ↓1 54.2
- 3 Claude Opus 5.5 51.6
- 4 Kimi K3.1 Pro 26.5
- 5
DeepSeek V4.1 Flash 19.0
- 6
Qwen 3.8 Max ↑1 15.0
- 7 GLM-5.3-Flash ↑2 12.7
- 8 MiMo-V2.6 Pro 12.0
- 9
Mistral Large 4 ↑1 10.4
- 10
MiniMax-M3.1-Flash ↓4 9.40
Placement
10 models we track have a score here.
- 1
Gemini 4 Argon ↑1 61.5
- 2 Claude Opus 5.5 ↑1 61.2
- 3 GPT-6 Astra ↓2 56.3
- 4 Kimi K3.1 Pro 33.4
- 5
DeepSeek V4.1 Flash 25.4
- 6
Qwen 3.8 Max ↑1 22.8
- 7
Mistral Large 4 ↑3 22.8
- 8
MiniMax-M3.1-Flash ↓2 21.8
- 9 GLM-5.3-Flash 20.0
- 10 MiMo-V2.6 Pro ↓2 16.6
Balance
10 models we track have a score here.
- 1
Gemini 4 Argon ↑1 59.8
- 2 Claude Opus 5.5 ↑1 58.0
- 3 GPT-6 Astra ↓2 55.4
- 4 Kimi K3.1 Pro 33.0
- 5
DeepSeek V4.1 Flash 28.4
- 6
MiniMax-M3.1-Flash 24.3
- 7
Mistral Large 4 ↑3 22.8
- 8
Qwen 3.8 Max ↓1 22.2
- 9 GLM-5.3-Flash 21.7
- 10 MiMo-V2.6 Pro ↓2 16.7
Response
10 models we track have a score here.
- 1 GPT-6 Astra 43.2
- 2 Claude Opus 5.5 ↑1 40.3
- 3
Gemini 4 Argon ↓1 39.4
- 4 Kimi K3.1 Pro 18.5
- 5
Qwen 3.8 Max ↑2 12.5
- 6
DeepSeek V4.1 Flash ↓1 10.9
- 7 GLM-5.3-Flash ↑2 10.4
- 8
MiniMax-M3.1-Flash ↓2 10.1
- 9
Mistral Large 4 ↑1 8.10
- 10 MiMo-V2.6 Pro ↓2 7.10
Affordance
10 models we track have a score here.
- 1
Gemini 4 Argon ↑1 38.1
- 2 Claude Opus 5.5 ↑1 37.3
- 3 GPT-6 Astra ↓2 35.7
- 4 Kimi K3.1 Pro 16.1
- 5
DeepSeek V4.1 Flash 10.9
- 6 MiMo-V2.6 Pro ↑2 10.7
- 7
MiniMax-M3.1-Flash ↓1 10.2
- 8
Qwen 3.8 Max ↓1 8.70
- 9 GLM-5.3-Flash 7.80
- 10
Mistral Large 4 7.60
Interaction
10 models we track have a score here.
- 1 Claude Opus 5.5 ↑2 33.9
- 2 GPT-6 Astra ↓1 31.2
- 3
Gemini 4 Argon ↓1 30.5
- 4
DeepSeek V4.1 Flash ↑1 20.4
- 5 Kimi K3.1 Pro ↓1 16.1
- 6 MiMo-V2.6 Pro ↑2 11.7
- 7 GLM-5.3-Flash ↑2 11.0
- 8
Qwen 3.8 Max ↓1 9.70
- 9
MiniMax-M3.1-Flash ↓3 7.20
- 10
Mistral Large 4 6.20
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official 3DSceneBench leaderboard
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official 3DSceneBench leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More 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%
Data audit
All 10 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access Public tasks, plus a private test set.
- Pass Confidence Checked against a primary source.
Scores 4 pass 1 note
- Pass Independent runs Scores from the official leaderboard (10 entries).
- Pass Models scored 10 models we track.
- Pass Newest score The newest is from October 2026, for Kimi K3.1 Pro.
- Pass Frontier models 3 of the top 10 have a score, led by GPT-6 Astra (#4).
- Note Headroom Scored as a plain number, with no fixed ceiling.
Leaderboard import 3 pass
- Pass Importer We read the board on pranalabs.github.io: 12 rows for models we track at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.