GPT-6 Astra scores highest on GoBench

A new benchmark called GoBench, designed to measure general reasoning abilities in LLMs through playing Go, shows GPT-6 Astra achieving the highest scores. GPT-6 Astra max reached 2568 Elo, surpassing Sol max (1929 Elo) and Opus 5 high (2076 Elo). Furthermore, Codex with Astra, when enabled for coding without internet access, achieved 3563 Elo, significantly outperforming Codex with Sol, which scored 2656 Elo.

This report introduces a new Go-playing benchmark for LLMs, showing GPT-6 Astra's superior performance compared to Sol max and Opus 5 high, unlike previous benchmarks.