GPT‑6.2 Astra and Sol

1302 points · 814 comments on HN · read original →

Points and comments are a snapshot, not live.

OpenAI releases GPT-6.2 Astra and Sol with better efficiency and lower cost per task.

OpenAI introduced GPT-6.2 Astra, a frontier model for complex work, and GPT-6.2 Sol, a more efficient model for everyday tasks. Astra sets new highs on Agents Last Exam (5.2 points above GPT-6 Astra at under a third the cost) and Terminal-Bench-Science. API prices: $8 input, $0.40 cached input, $40 output per million tokens for Astra; Sol retains GPT-6.1 Sol's $2/$0.10/$10 rates. Both models improve on coding, document creation, and computer use. Safety precautions include controls against dual-use capabilities in cybersecurity, biology, and AI self-improvement.

Available in ChatGPT and the API. Early testers report better slide design, spreadsheet work, and code quality with fewer revisions.

What commenters are saying

Commenters noted the apparent timing of the release just before Anthropic's IPO, with several seeing it as strategic. Many praised the focus on lower cost per task rather than raw performance. A significant thread discussed the tradeoffs between Astra and Sol: some found Astra noticeably better for complex coding, subagent coordination, and design work, while others saw little difference in everyday use. Several users criticized Luna model laziness and decreasing reliability. Technical complaints included the announcement page being laggy, crashing browsers, and pegging GPUs, though one user reported it ran well on Linux.