Bug Hunt Bench: GPT-6.2 Astra Max fixes 62.7/105 planted bugs, 20 remain unsolved by any model

PawelHuryn · x · 2026-11-19

Pawel Huryn published full effort-level results from Bug Hunt Bench (planted bugs only, 105 total). GPT-6.2 Astra MAX leads with 62.7/105; across hundreds of runs, 20/105 bugs have never been solved by any frontier model. Anthropic's Fable 5.2 MAX scores 58/105 at a steep $139.65 per run, while Opus 5.5 MAX manages just 41.7/105 — highlighting a persistent gap between the best coding models and complete bug coverage.

Related event: Google Launches Gemini 4 Neon, Topping Coding Agent Benchmarks at Low Cost(167 posts)→

Original post →

More from coding & agent

coding & agent channel →