Agentic sandbox WorldBox doubles GPT-6 Astra's MazeBench score, 23 to 61 gems
ZhitingHu · x · 2026-10-16
EnactraAI unveiled WorldBox, an agentic sandbox for spatial reasoning. On MazeBench — a 256-room long-horizon spatial reasoning test where even Opus 5.5 burned 2.2B tokens for just 13/100 gems — GPT-6 Astra scored 23 gems alone but 61 with WorldBox attached. The sandbox handles long-term memory, inferring complex mechanics, and simulated reasoning.
More from coding & agent
- LangChain answers the most common questions about agent evals: datasets, judges, and trajectories — LangChain · 2026-10-16
- Signal65 benchmarks: Intel Xeon 6 runs agent workloads 11% faster for less money — ryanshrout · 2026-10-16
- Vercel Daemons in production: log delivery pipeline hits 1 Gbit per second and ramping — cramforce · 2026-10-16
- From Mathematica to ~1000 Python scripts: a 20-year math visualization journey — S_Conradi · 2026-10-16
- How to Get Better at Delegating to Your AI Assistant: It's a Learnable Skill — craigsdennis · 2026-10-16
- HeyGen MCP turns ChatGPT recipes into personalized cooking videos — toolstelegraph · 2026-10-16