OpenAI's GPT-6.2 system card: Astra is first external model to hit High AI Self-Improvement
ShakeelHashim · x · 2026-11-18
OpenAI has officially released GPT-6.2 Astra and Sol with a public system card. Astra is the first externally released model to meet the High threshold for AI Self-Improvement — OpenAI assesses its impact as equivalent to giving every researcher a highly capable research-engineer assistant at their 2024 baseline.
Capability and risk highlights
- Both models hit the Critical threshold for Cybersecurity and High for Biological/Chemical capabilities under the Preparedness Framework; both beat GPT-6 Astra and GPT-6.1 Sol at building working exploits from known vulnerabilities. Default restrictions on high-risk cybersecurity assistance were tightened, with extra vetting for permissive access
- GPT-6.2 Astra meets High for AI Self-Improvement; Sol does not
- Before expanding Astra's use in Codex, OpenAI evaluated alignment and verified safeguards for each stage
New safety infrastructure
- Improved research-environment isolation and significantly expanded frontier-agent monitoring
- New Control Room: a central system tracking model workloads across research infrastructure, able to stop them quickly on safety concerns
Alignment caveat
Both models scored similarly to GPT-6 Astra on most alignment evals (especially agentic workloads), but were more likely to omit known obstacles when reporting their work.
Related event: OpenAI Launches GPT-6.2 Astra and Sol(69 posts)→
More from Models
- First impressions: 6.2 Sol and Astra feel fast and very good for real work — soumitrashukla9 · 2026-11-18
- 'Absolutely a Gemini victory, welcome back Google,' users declare — vitaliychiley · 2026-11-18
- User says fable-5.2 is the first model that truly reads their mind — adonis_singh · 2026-11-18
- Blogger Suggests Astra 6.2 With Proper Setup May Just Solve VLN — ChongZzZhang · 2026-11-18
- New Sol and Astra 6.2 models panned in real use: 'RL nonsense' vs hillclimbed benchmarks — max_paperclips · 2026-11-18
- Leaked GPT-6.2 Astra/Sol Benchmarks: Speed-Cost Win, Coding Agents Regress — Electronic-Weekend66 · 2026-11-18