GPT-5.6 Sol: how it does on security work, and what it will answer

What happened

$4 and $20 per million tokens. RealVuln scores it F3 74.7 across all 140 repositories at 59.33% precision and 76.92% recall, 4.8 points behind Daybreak Blue on identical weights, and OpenAI's own table gives 78.5% on ExploitBench. In Codex a refused turn retries once on the Daybreak model without changing the session's stored model.

RealVuln 3.1.0, F3 (micro): 74.7 (independent; Codex CLI (gpt-5.6-sol-codex-cli), prompt sha256:45a1200d61e6 (tsjs-v1); 140 of 140 repositories pinned by commit SHA (full corpus, TypeScript and JavaScript included); safeguards: production safeguards as deployed on the API; run 2026-09-11; CONFIRMED). strict F3 74.7; precision 59.33%, recall 76.92%; $560.34 for the run; $0.0756 per 100 lines; 568.6s wall clock a run; reasoning effort high. ExploitBench, score: 78.5% (vendor-card; OpenAI comparison table; ExploitBench as run by OpenAI; safeguards: safeguard state unstated in the summaries read; run 2026-09-03; SINGLE). Refusal gate (CONFIRMED, observed 2026-09-11): Refused turns retry once on the configured Daybreak model; only a refused turn routes there.