OpenAI GPT-6.2 Astra and Sol System CardProvider report
2026-11-17Seven capability metrics from the GPT-6.2 Astra and Sol system card, with same-source comparisons for GPT-6 Sol, GPT-6 Astra, GPT-6.1 Sol, and GPT-6.1 Terra. HealthBench Professional uses length-adjusted scores, retaining unadjusted scores and response lengths; OpenAI no longer reports the original HealthBench for new models, so it has no new scores here. Four MentalHealthBench metrics use maximum reasoning effort with per-model standard errors and task counts. Three cybersecurity rows come from Table 4: SEC-Bench Pro is pass@1, with maximum effort stated for the new models; ExploitBench V8 is mean capability credit, separate from arbitrary code-execution rate, and historical vulnerabilities may cause contamination. Comparison models may reflect later versions; historical release snapshots are unchanged.