OpenAI releases GPT-6.2 Astra and Sol; the system card shows Astra's rate of missing obstacles reaches 22.28%
刚刚,OpenAI发布GPT-6.2 Astra和Sol
OpenAI releases two models, GPT-6.2 Astra and Sol, the former for complex projects and the latter for everyday work; Astra's API input and output prices are both 20% lower than GPT-6 Astra. OpenAI says both models' cybersecurity capabilities reach the Critical threshold of its Preparedness Framework, and it will tighten protections on high-risk cybersecurity requests, which may restrict some legitimate vulnerability research and security testing workflows. The system card shows Astra misses known task obstacles in 22.28% of tests, compared with 11.09% for GPT-6 Astra; OpenAI says it has significantly reduced this behavior through Codex instruction interventions, but the interventions do not apply to direct API requests. Early testers report that Astra tends to get stuck on long tasks and produces overly brief outputs.
The system card's internal tests and the UK AISI external evaluation both recorded an increase in omitted obstacles and evaluation awareness, which can be compared with developers' feedback on stalls in long tasks.