GPT-6.2 System Card (GPT-6.2 Astra and GPT-6.2 Sol)
System card for OpenAI's GPT-6.2 Astra and GPT-6.2 Sol models, published on the Deployment Safety Hub on 17 November 2026. Alongside the preparedness and safeguards chapters, it reports safe-completion evaluations including users under 18 (eating disorders, emotional reliance, self-harm), dynamic adversarial-user simulations for self-harm and emotional reliance, HealthBench Professional, and MentalHealthBench, with comparisons to GPT-6 Astra, GPT-6.1 Sol and earlier releases.
Publisher
OpenAI
Published
17 Nov 2026
Added
2 days ago
DOI
—
Key Findings
- MentalHealthBench overall (1,215 synthetic conversations, four responses per task, maximum reasoning effort) is 53.3% for GPT-6.2 Astra and 50.9% for GPT-6.2 Sol, against 58.7% for GPT-6 Astra and 57.9% for GPT-6.1 Sol; both new models score lower at every acuity level (emergent subset: 51.0% Astra, 52.6% Sol, against 58.5% and 58.0%).
- The card attributes about half of the MentalHealthBench decline to the 'context seeking and assessment' axis and links it to a stated preference for concise, actionable replies; it adds that the reduced context seeking is broader than mental-health conversations.
- In dynamic adversarial simulations (50 conversations per category and model), the mean share of compliant replies per conversation on self-harm is 91.7% for GPT-6.2 Astra and 95.1% for GPT-6.2 Sol, against 98.9% for GPT-6 Astra; on emotional reliance the figures are 98.9% (Astra) and 99.7% (Sol).
- On the under-18 safety evaluations with teen protections enabled, GPT-6.2 Astra scores 93.1% on eating disorders, 96.6% on emotional reliance and 99.7% on self-harm; GPT-6.2 Sol scores 92.7%, 95.6% and 99.7%. The card states that in manual review the new models are 'more willing to be helpful, occasionally prioritizing helpfulness over safety on unsafe prompts'.
- HealthBench Professional length-adjusted scores are 66.3 (Astra) and 65.9 (Sol) against 64.7 for GPT-6 Astra; unadjusted scores are 66.0 and 65.5 against 68.2, with mean responses of about 1,880 characters against 3,185.
Methodology Notes
Vendor self-report; no external replication. Values were read from the HTML card on the OpenAI Deployment Safety Hub (page text, Tables 14, 17 to 19 and the section 3.4.2 narrative; published 17 November 2026). The card states that evaluation scenarios were selected for difficulty and that scores do not estimate violation rates in ordinary use; safe-completion evaluations are run without system-level safeguards. The dynamic-simulation per-category table was read from the narrative (Table 18 cells were not machine-readable); the narrative gives Astra, Sol and GPT-6 Astra values in that order. Evaluation versions and baselines change between cards, so figures should not be compared with other cards.
Sources
Topics
Tags
Cite This
APA
OpenAI. (2026). GPT-6.2 System Card (GPT-6.2 Astra and GPT-6.2 Sol). https://deploymentsafety.openai.com/gpt-6-2-astra-sol