Code4Scene
Agents build and edit Unreal Engine scenes. The evaluator inspects the scene itself for correct edits, intact assets, and physical validity.
- Made by
- Xiaokang Ye and colleagues
- Released
- September 2026
- Kind
- Image generation and editing
- Level
- University
- Scored
- Checked automatically
- Answers
- Code
- Tasks
- Public, plus a private test set
- Measures
- Capability
- Language
- English
- Interaction
- Agentic
- Environment
- Unreal Engine
- Refresh
- One fixed set
- Run by
- A host
- In views
- Agents, Vision, Image, video and 3D generation
- Strongest model tested
- GPT-6 Astra #4 on the capability index
- Last model added
- September 2026
- Scores refreshed
- November 20th, 2026
GPT-6 Astra leads with 0.62 among models released by September 2026
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. 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:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- Claude Opus 5.5
- and 12 more
- 1 GPT-6 Astra max 0.62
- 2
Gemini 3.8 Flash high 0.62
- 3 Claude Fable 5.1 max 0.61
- 4 Claude Opus 5 max 0.59
- 5 GPT-5.6 Sol high 0.55
- 6
Muse Spark 1.3 medium 0.50
- 7 GLM-5.3-Flash max 0.41
- 8
Qwen 3.8 27B thinking off 0.38
- 9 Grok 4.6 high 0.38
- 10
Inkling high 0.31
- 11
Gemma 4 31B IT thinking on 0.30
- 12
DeepSeek V4.1 Flash high 0.23
Source: Official Code4Scene leaderboard
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-6 Astra | 0.62 | max | Sep 2026 | Official leaderboard |
| 2 | | 0.62 | high | Sep 2026 | Official leaderboard |
| 3 | Claude Fable 5.1 | 0.61 | max | Sep 2026 | Official leaderboard |
| 4 | Claude Opus 5 | 0.59 | max | Jul 2026 | Official leaderboard |
| 5 | GPT-5.6 Sol | 0.55 | high | Jul 2026 | Official leaderboard |
| 6 | | 0.50 | medium | Sep 2026 | Official leaderboard |
| 7 | GLM-5.3-Flash | 0.41 | max | Aug 2026 | Official leaderboard |
| 8 | | 0.38 | thinking off | Aug 2026 | Official leaderboard |
| 9 | Grok 4.6 | 0.38 | high | Aug 2026 | Official leaderboard |
| 10 | | 0.31 | high | Jul 2026 | Official leaderboard |
| 11 | | 0.30 | thinking on | Apr 2026 | Official leaderboard |
| 12 | | 0.23 | high | Sep 2026 | Official leaderboard |
Rankings by part
We read 6 parts from the benchmark’s own leaderboard, each ranking models on one part of the test, such as Text-to-Scene, Image-to-Scene and Repair F1. Gemini 3.8 Flash leads 3 of them, and 3 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
Text-to-Scene
12 models we track have a score here.
- 1 Claude Fable 5.1 ↑2 0.79
- 2 GPT-6 Astra ↓1 0.72
- 3 Claude Opus 5 ↑1 0.72
- 4 GPT-5.6 Sol ↑1 0.71
- 5
Gemini 3.8 Flash ↓3 0.66
- 6
Muse Spark 1.3 0.65
- 7 Grok 4.6 ↑2 0.57
- 8
Qwen 3.8 27B 0.56
- 9 GLM-5.3-Flash ↓2 0.51
- 10
Gemma 4 31B IT ↑1 0.47
Image-to-Scene
12 models we track have a score here.
- 1
Gemini 3.8 Flash ↑1 0.58
- 2 GPT-6 Astra ↓1 0.51
- 3 Claude Opus 5 ↑1 0.47
- 4 Claude Fable 5.1 ↓1 0.42
- 5 GPT-5.6 Sol 0.39
- 6
Muse Spark 1.3 0.36
- 7 GLM-5.3-Flash 0.32
- 8
Qwen 3.8 27B 0.23
- 9
DeepSeek V4.1 Flash ↑3 0.21
- 10
Inkling 0.20
Repair F1
12 models we track have a score here.
- 1
Gemini 3.8 Flash ↑1 0.53
- 2 GPT-6 Astra ↓1 0.45
- 3 Claude Opus 5 ↑1 0.39
- 4 Claude Fable 5.1 ↓1 0.33
- 5 GPT-5.6 Sol 0.32
- 6
Muse Spark 1.3 0.25
- 7 GLM-5.3-Flash 0.19
- 8
Qwen 3.8 27B 0.12
- 9
Inkling ↑1 0.10
- 10
DeepSeek V4.1 Flash ↑2 0.07
Image-to-Scene physical safety
12 models we track have a score here.
- 1 GLM-5.3-Flash ↑6 0.85
- 2
Muse Spark 1.3 ↑4 0.81
- 3
Gemini 3.8 Flash ↓1 0.80
- 4 GPT-6 Astra ↓3 0.80
- 5 Claude Fable 5.1 ↓2 0.80
- 6
DeepSeek V4.1 Flash ↑6 0.79
- 7 Claude Opus 5 ↓3 0.79
- 8 Grok 4.6 ↑1 0.74
- 9 GPT-5.6 Sol ↓4 0.69
- 10
Qwen 3.8 27B ↓2 0.67
Indoor
12 models we track have a score here.
- 1 GPT-6 Astra 0.79
- 2
Gemini 3.8 Flash 0.72
- 3
Muse Spark 1.3 ↑3 0.63
- 4 Claude Opus 5 0.62
- 5 Claude Fable 5.1 ↓2 0.54
- 6 GPT-5.6 Sol ↓1 0.50
- 7 GLM-5.3-Flash 0.47
- 8
Qwen 3.8 27B 0.30
- 9 Grok 4.6 0.20
- 10
DeepSeek V4.1 Flash ↑2 0.19
Outdoor
12 models we track have a score here.
- 1
Gemini 3.8 Flash ↑1 0.51
- 2 Claude Opus 5 ↑2 0.39
- 3 GPT-6 Astra ↓2 0.38
- 4 Claude Fable 5.1 ↓1 0.37
- 5 GPT-5.6 Sol 0.34
- 6 GLM-5.3-Flash ↑1 0.25
- 7
DeepSeek V4.1 Flash ↑5 0.22
- 8
Muse Spark 1.3 ↓2 0.22
- 9
Inkling ↑1 0.21
- 10
Qwen 3.8 27B ↓2 0.19
The best score, 0.62, costs $13 a task
Runs on the cost frontier score higher than any run that's cheaper. Gemini 3.8 Flash at high comes within 0 of the best for $1.92 a task, 7 times cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.
Source: Official Code4Scene leaderboard
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| GPT-6 Astra | max | 0.62 | $13 | ||
| high | 0.62 | $1.92 | |||
| Claude Fable 5.1 | max | 0.61 | $11 | ||
| Claude Opus 5 | max | 0.59 | $19 | ||
| GPT-5.6 Sol | high | 0.55 | $2.98 | ||
| medium | 0.50 | $0.09 | |||
| GLM-5.3-Flash | max | 0.41 | $0.40 | ||
| thinking off | 0.38 | $0.38 | |||
| Grok 4.6 | high | 0.38 | $2.29 | ||
| thinking on | 0.37 | $0.30 | |||
| high | 0.31 | $0.48 | |||
| thinking on | 0.30 | $0.03 | |||
| thinking off | 0.28 | $0.02 | |||
| high | 0.23 | $0.12 |
3 models have held number one on the official leaderboard since its release in September 2026
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official Code4Scene 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 Code4Scene 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 Vision
- MMMU In use 11.5K college-level multimodal questions across 30 subjects. Multiple-choice and open accuracy.
Qwen 3.6 Plus 86%
- MMMU-Pro In use Harder MMMU: college-level multimodal questions with more options and vision-only inputs to remove text shortcuts. GPT-5.6 Sol 84.6%
- CharXiv In use Chart understanding on real arXiv figures. Reasoning questions require synthesizing information from complex charts. Claude Fable 5 93.5%
- MathVista In use Math reasoning in visual contexts (charts, geometry, puzzles) from 31 datasets. Answer accuracy.
Seed2.1 Pro 90.7%
- MATH-Vision (MathVision) In use 3,040 competition math problems with visual context. Answer accuracy.
Qwen 3.8 Max 97.7%
- MMStar In use Models answer 1,500 human-curated image questions that cannot be solved without the image, across 6 core capabilities. Scored by accuracy.
Qwen 3.8 Max 85.9%
Data audit
All 9 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 3 pass 2 notes
- Pass Independent runs Scores from the official leaderboard (12 entries).
- Pass Models scored 12 models we track.
- Pass Newest score The newest is from September 2026, for DeepSeek V4.1 Flash.
- Note Frontier models Only 2 of the top 10 have a score. The strongest tested is 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 simworld-ai-code4scene.static.hf.space: 14 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.