KradleBench benchmarks frontier AI in multi-agent Minecraft: GPT-6.2 Astra scores 78.3, beats all but one human
Vjeux · x · 2026-11-19
Kradle launched KradleBench, a benchmark set in messy, dynamic multi-agent Minecraft worlds that's hard to saturate, with full traces, logs and 3D interactive replays for every run.
Key findings
- Overall: GPT-6.2 Astra tops at 78.3, skilled human SethBling second at 77.2, GPT-6.2 Sol 67.9; Claude Opus 5.5 only 39.1, Gemini 4 Carbon #16 due to latency.
- OpenAI models dominate because they iterate and adapt faster: Astra iterates 4x faster than Fable 5.2, while Claude models think longer then stick to their strategy.
- Across five dimensions (Collaboration, Economy, Conflict, Logic, Creativity), Astra leads collaboration/logic/creativity, but SethBling still beats all models at Economy (100) and Conflict (90); top models win solo logic and creativity challenges.
- Models can coordinate complex tasks and build recognizable objects (Godzilla, Golden Gate Bridge), though quality gaps are large.
Related event: KradleBench Launches: Astra 6.2 Dethrones Minecraft Legend SethBling(6 posts)→
More from Models
- Resolve launches SRE-1, a 27B RL-trained model for root-causing production incidents — tianyin_xu · 2026-11-19
- Theo does a about-face: 'Gemini 4 Neon might actually be a good model' — vitaliychiley · 2026-11-19
- Artificial Analysis Launches AA-PersonalityBench: 30 Traits, 8 Indexes, 53 Models — ArtificialAnlys · 2026-11-19
- GPT-6.2 writes 40% shorter, less structured replies and admits uncertainty far less, finds Artificial Analysis — ArtificialAnlys · 2026-11-19
- DeepSeek 4.1-Flash slips into game roles easily while Nemotron models need heavy coaxing — rickasaurus · 2026-11-19
- Leaked "Claude Mythos 5.5" strings spotted on Anthropic's site ahead of possible launch — imjustnewatai · 2026-11-19