Skip to content

Benchmark

RecreationBench paper · macOS · Visual assertion score

Scores are percentages. Higher is better.

Models and systems measured
10
Scores recorded
10
Sources
1
IQ weight
1.23

Score distribution

24.2%
Mean
25.8%
Median
9.8%
Std dev
13%
10th percentile
34.5%
90th percentile
5.9%
Min
5.9%42.7%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 1.23 for chat models, measured over 8 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#Model or systemScore
1GPT-6 AstraRecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 272k context · 128k max output tokens · macOS · nominal 50-task split · platform-specific score· Reported by QwenLM / RecreationWorld authors· via BenchBoard docs42.7%
2GPT-5.6 SolRecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 272k context · 128k max output tokens · macOS · nominal 50-task split · platform-specific score· Reported by QwenLM / RecreationWorld authors· via BenchBoard docs34.5%
3Claude Opus 5RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · macOS · nominal 50-task split · platform-specific score· Reported by QwenLM / RecreationWorld authors· via BenchBoard docs26.4%
4Claude Opus 4.8RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · macOS · nominal 50-task split · platform-specific score· Reported by QwenLM / RecreationWorld authors· via BenchBoard docs25.8%
5GLM 5.3RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · macOS · nominal 50-task split · platform-specific score· Reported by QwenLM / RecreationWorld authors· via BenchBoard docs20.9%
6Kimi K3RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · max reasoning effort per site data · 1m context · 128k max output tokens · macOS · nominal 50-task split · platform-specific score· Reported by QwenLM / RecreationWorld authors· via BenchBoard docs19.9%
7Gemini 3.7 FlashRecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · high reasoning effort per site data · 1m context · 64k max output tokens · macOS · nominal 50-task split · platform-specific score· Reported by QwenLM / RecreationWorld authors· via BenchBoard docs13%
8Qwen3.7 PlusRecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · thinking enabled per site data · 256k context · 32k max output tokens · macOS · nominal 50-task split · platform-specific score· Reported by QwenLM / RecreationWorld authors· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.5.9%

Not ranked

The source marks these results as descriptive, not capability scores, so we list them by name instead of ranking them.

Model or systemScore
Grok 4.6RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · xhigh reasoning effort per site data · 500k context · 128k max output tokens · macOS · nominal 50-task split · platform-specific score· Reported by QwenLM / RecreationWorld authors· via BenchBoard docs24.1%
Qwen3.8 Max 0902RecreationWorld benchmark-owner paper run · 250 held-out tasks · GUI exploration + coding tools + visual verification · xhigh reasoning effort per site data · 256k context · 32k max output tokens · macOS · nominal 50-task split · platform-specific score· Reported by QwenLM / RecreationWorld authors· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.29.2%

Data credits: BenchBoard by SYuan03 (github.com/SYuan03/benchboard). MIT covers its code and data organization; no results licence is inferred. Numeric relays retain the original stating-source links.