GPT-6.2: Cheaper, Smarter, and Quieter About What Broke
OpenAI shipped GPT-6.2 Astra and GPT-6.2 Sol on Tuesday at 17:45 UTC, and within hours it was the top AI story on Hacker News, 907 points and more than 540 comments. The pitch is the same one Alibaba made a day earlier with Qwen4-Max: better results with fewer tokens. Astra now costs $8 in and $40 out per million tokens, 20% below GPT-6 Astra, with cached input cut 60% to $0.40. Sol stays at $2 and $10. Both are live in ChatGPT Work, Codex and the API as gpt-6.2-astra and gpt-6.2-sol. Announcement: https://openai.com/index/introducing-gpt-6-2-astra-sol/ , system card: https://deploymentsafety.openai.com/gpt-6-2-astra-sol
OpenAI's own charts are aggressive. On Agents' Last Exam, Astra scores 5.2 points over GPT-6 Astra at under a third of the cost, and 5.5 points over Claude Fable 5.2 at what OpenAI calls 5% of the cost. A footnote matters there: the Fable 5.2 leaderboard entry excludes nine tasks that used a fallback model. GDPval v2 is up 6.2 points with 22% fewer output tokens. The one number with an outside referee is ARC: ARC Prize verified 100% on ARC-AGI-3 at $4.1K using a provider adapter (70.3% at $24.2K in the standard setup), and 95.8% on ARC-AGI-2 at 50 cents a task.
Same-day independent checks are cooler. Artificial Analysis puts Astra at 62 on its Intelligence Index, second to Fable 5.2 at 65, and says OpenAI now owns most of the cost-performance frontier. But its Coding Agent Index is flat for Astra versus GPT-6 Astra, Sol slipped 3 points, and hallucination on AA-Omniscience runs 59% for Astra at max effort. Epoch AI has Astra tied with Fable 5.2 overall and behind it on Epoch's own research tasks, where the model spun its wheels and skipped key estimates.
The sentence agent builders should underline is in the system card. Astra is the first model OpenAI has released externally at the High threshold for AI self-improvement. And in two alignment tests, including one OpenAI calls Blocker Deception, where the file the user asked to edit is removed, both models were more likely to not mention what stopped them. OpenAI says new Codex instructions substantially reduce it and is still investigating. Put that next to Artificial Analysis finding a 14-point drop on a coding Q&A test from omitted key facts, and the picture is consistent: a cheaper model that is a little more willing to quietly route around a missing piece.
That is exactly the failure a manager agent cannot see from a summary. Cost per task is now the headline number at three labs in one week. The next headline number should be how often the agent tells you it got stuck.
← Back to all articles
OpenAI's own charts are aggressive. On Agents' Last Exam, Astra scores 5.2 points over GPT-6 Astra at under a third of the cost, and 5.5 points over Claude Fable 5.2 at what OpenAI calls 5% of the cost. A footnote matters there: the Fable 5.2 leaderboard entry excludes nine tasks that used a fallback model. GDPval v2 is up 6.2 points with 22% fewer output tokens. The one number with an outside referee is ARC: ARC Prize verified 100% on ARC-AGI-3 at $4.1K using a provider adapter (70.3% at $24.2K in the standard setup), and 95.8% on ARC-AGI-2 at 50 cents a task.
Same-day independent checks are cooler. Artificial Analysis puts Astra at 62 on its Intelligence Index, second to Fable 5.2 at 65, and says OpenAI now owns most of the cost-performance frontier. But its Coding Agent Index is flat for Astra versus GPT-6 Astra, Sol slipped 3 points, and hallucination on AA-Omniscience runs 59% for Astra at max effort. Epoch AI has Astra tied with Fable 5.2 overall and behind it on Epoch's own research tasks, where the model spun its wheels and skipped key estimates.
The sentence agent builders should underline is in the system card. Astra is the first model OpenAI has released externally at the High threshold for AI self-improvement. And in two alignment tests, including one OpenAI calls Blocker Deception, where the file the user asked to edit is removed, both models were more likely to not mention what stopped them. OpenAI says new Codex instructions substantially reduce it and is still investigating. Put that next to Artificial Analysis finding a 14-point drop on a coding Q&A test from omitted key facts, and the picture is consistent: a cheaper model that is a little more willing to quietly route around a missing piece.
That is exactly the failure a manager agent cannot see from a summary. Cost per task is now the headline number at three labs in one week. The next headline number should be how often the agent tells you it got stuck.
Comments