New Minecraft benchmark KradleBench: GPT-6.2 Astra edges a skilled human 78.3 vs 77.2 across 13 challenges
TheMrZZ0 · reddit · 2026-11-20
The KradleBench team released a Minecraft AI benchmark: 26 models, 13 challenges, 345 games per model. GPT-6.2 Astra topped the board at 78.3/100, just above skilled human player SethBling's 77.2; the next model, GPT-6.2 Sol, scored 67.9.
How it works: agents receive text observations (position, health, inventory, nearby blocks, chat) and act by writing JavaScript. Five weighted dimensions: Collaboration 30%, Strategy 30%, Conflict 30%, Logic (redstone) 5%, Creativity 5%.
Standout findings:
- Human and top model win differently: SethBling stronger at Strategy and Conflict; Astra at Collaboration and Logic.
- In combat the human exploited AI thinking latency — the observe→call→run-code loop takes seconds, and SethBling killed AIs mid-loot or mid-plan.
- Astra used an unprompted capability: building "a pelican riding a bicycle," it summoned rotated block display entities to make round, tilted wheels, bypassing Minecraft's block grid, in 4 of 5 builds. No other model tried.
- Best models coordinate over chat: three Astras split by material then swapped duplicates to cover 106 of 108 inventory slots in Scavenger Hunt.
- Price correlates loosely with score: Sol cost a quarter of Astra for 67.9; Claude Fable 5.2 cost 7% more than Astra but scored only 37.8.
Full leaderboard, per-challenge results and replays at kradle.ai.
Related event: KradleBench Launch: GPT-6.2 Astra Edges Out Minecraft Legend SethBling(8 posts)→
More from Fun
- Gemini 4 Neon reportedly cheers on its subagents' results mid-task — sebkrier · 2026-11-20
- AI agent phones a roof repair company, leak fixed within an hour for $300 — stuffyokodraws · 2026-11-20
- Snap a photo on your phone and drop it straight into VR: "spatial airdropping" demo — dreamwieber · 2026-11-20
- Apple Puts a Daily Limit on Siri and the Internet Is Roasting It — alfcnz · 2026-11-20
- Interns Present AI Token-Cost Projects at 'From Lab to Launch' Fall Intern Science Fair — floguo · 2026-11-20
- ML author Burkov slams AI-written love notes: 'seriously ill' — burkov · 2026-11-20