Benchmarks
357 sources · 1,263 models × 3,079 benchmarks · 162,127 scoresKradle's KradleBench puts models in live, multi-agent Minecraft worlds, where each model reads JSON observations and acts by writing JavaScript against a fixed skills library while teammates, opponents, creatures and the clock keep moving. Its overall score out of 100 weights five dimensions: Collaboration, Economy and Conflict at 30% each (team builds, a crop market and battle-royale and survival games against copies of reference models) and Logic and Creativity at 5% each (redstone circuits and judged open-brief builds), over 345 runs of 13 challenges per model in battery kb-1.8.2. The launch board also scores a human player, SethBling, at 77.21, second to GPT-6.2 Astra's 78.28; human rows are not imported.
633rd strongest of 818 fitted benchmarks
Signal-to-noise +3.32
83rd hardest of 818 fitted benchmarks
Capability needed ≳ 172.9
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | GPT-6.2 Astra | medium | 78.3 | |
| 2 | GPT-6.2 Sol | medium | 67.9 | |
| 3 | GPT-6 Astra | not reported | 56.1 | |
| 4 | GPT-6.1 Sol | not reported | 54.6 | |
| 5 | GPT-6.1 Terra | not reported | 48.6 | |
| 6 | Gemini 4 Hydrogen | low | 43.9 | |
| 7 | Claude Opus 5.5 | not reported | 39.1 | |
| 8 | Claude Fable 5.2 | low | 37.8 | |
| 9 | Claude Sonnet 5.5 | not reported | 33.8 | |
| 10 | Gemini 4 Argon | low | 28.6 | |
| 11 | Muse Wonder | low | 27.6 | |
| 12 | Muse Ember | not reported | 25.7 | |
| 13 | Gemini 3.8 Flash | low | 24.3 | |
| 14 | Grok 4.7 | low | 23.0 | |
| 15 | Gemini 4 Carbon | low | 22.0 | |
| 16 | Muse Spark 1.3 | not reported | 21.7 | |
| 17 | GLM-5.3 | low | 21.2 | |
| 18 | GPT-6 Luna | not reported | 18.1 | |
| 19 | Gemma 4 31B IT | disabled | 17.1 | |
| 20 | Kimi K3.1 Pro | — | 15.5 | |
| 21 | Muse Spark Open | not reported | 15.2 | |
| 22 | Claude Haiku 5.5 | not reported | 15.0 | |
| 23 | Nemotron 3.5 Super VL 120B A12B (precision unspecified) | disabled | 11.2 | |
| 24 | Qwen4-Max | low | 11.0 |
Benchmarks outside its family whose model rankings track it most reliably (Spearman ρ over at least 10 shared models).