MazeBench record: Schema with GPT-6 Astra hits 33%, beating top-50 human median
patience_cave · x · 2026-10-14
A new record on the long-horizon MazeBench benchmark: the Schema harness with GPT-6 Astra scored 33%, up from 23% with Codex.
- The leaderboard shows gpt-6-astra + Schema at 33%, the same model on Codex max at 23%, gpt-6.1-sol at 17%, and claude-opus-5-5 at 13%
- Schema with GPT-6 Astra collected 33 gems and visited 139 rooms, exceeding the top-50 human median on both metrics
- Key insight: its program retains and reuses learned rules across context compactions — a concrete demonstration of persistent memory for long-horizon agents
Work by @guanningzeng and @HavenFeng.
Related event: Schema with GPT-6 Astra Sets New MazeBench Record at 33%(2 posts)→
More from coding & agent
- RoboSimStudio: Claude and GPT build robot digital twins from phone video in an hour — Scobleizer · 2026-10-15
- FactoryAI launches OpenCompute, a self-hosted orchestrator for agent sandboxes in your VPC — annbordetsky · 2026-10-15
- shadcn praises Claude Code's clever 'Pin as chapter' feature — shadcn · 2026-10-15
- Team ships four developer guides for Gemini 4 Argon, incl. 1M-token long decode — patloeber · 2026-10-15
- Dev Creates Anthropic-Inspired Color Themes for IntelliJ — therealdanvega · 2026-10-15
- uv team split one giant crate into 14, cutting incremental Rust build times by 30-43% — charliermarsh · 2026-10-15