CodeActionBenchCode-as-Policy Manipulation
← Back to the benchmarkObjective head-to-head

Same task, direct verifier comparison

The two rails align by task. Every node reports the three accepted repeat outcomes without inferring stages, intent, or recovery quality.

GPT-6 Astra (Codex)55/75successful episodes
ASR delta?ASR delta is always A minus B: positive means A is higher; negative means A is lower.+28.00 pp

11 A leads · 11 Tied · 3 B leads

Opus 5 (Claude Code)34/75successful episodes
GPT-6 Astra (Codex)25 tasksOpus 5 (Claude Code)