TRACES benchmark launches: GPT-6 Astra tops charts but lags in self-repair

New evaluator TRACES launched a six-dimension benchmark for agentic capabilities in hard scientific discovery tasks. GPT-6 Astra topped the leaderboard with gains across all dimensions, but self-repair improved the least and still lags Opus-5.

2026-09-12 ~ 2026-09-13 · 2 related posts

• TRACES Board: GPT-6-astra Tops Agent Leaderboard but Trails Opus-5 on Self-Repair — thisdudelikesAI · 2026-09-12
• TRACES benchmark: GPT-6 Astra gains on all six axes but Repair lags — rohanpaul_ai · 2026-09-13