OpenAI ships GPT-6.2 Astra, first deployed model to trip its own self-improvement alarm
OpenAI dropped GPT-6.2 Astra on November 17, 2026, and it's the first externally deployed OpenAI model to hit the High threshold for AI self-improvement under the company's Preparedness Framework. Its cheaper sibling, GPT-6.2 Sol, landed at Critical for cybersecurity. Both are live in ChatGPT Work, Codex and the API, and Astra costs less than the model it replaces.
- Original source: cryptobriefing.com.
- The source published it on Nov 17, 2026 at 6:53 PM.
- No second outlet has reported it yet. Treat it as a single-source claim.
OpenAI rolled out two models at once on November 17, 2026: GPT-6.2 Astra and GPT-6.2 Sol. Both are already live in ChatGPT Work, in Codex and through the API. OpenAI frames the pair as a split lineup — Astra is the heavier, more capable flagship, while Sol is the cost-effective option built for everyday use. Both target professional workloads, and the company calls them more efficient than what came before.
The safety labels are the headline. Astra is the first externally deployed OpenAI model to meet the High threshold for AI Self-Improvement under the company's Preparedness Framework — the scorecard OpenAI built to track exactly this capability. Sol didn't reach High on self-improvement, but it hits the Critical threshold for cybersecurity and the High threshold for biological and chemical risks. Same release, two very different warning labels.
Astra takes over from GPT-6 Astra, which OpenAI shipped in September 2026. That earlier model posted a perfect score on ExploitBench but still fell short of the High self-improvement bar. Roughly two months later, its successor clears it. On the benchmarking side, GPT-6.2 Astra improves on GPT-6 Astra by 5.2 points on Agents' Last Exam.
OpenAI also cut the price. GPT-6.2 Astra costs less than GPT-6 Astra, with general-user pricing set at $8/$40 per million tokens. The two-tier structure lets teams route demanding work to Astra and routine tasks to Sol, and shipping both through ChatGPT Work, Codex and the API on day one puts them in front of business users, coders and developers immediately.
Self-improvement worries about something different from the usual AI risk categories, which describe what a model could help a person do — assist with a cyberattack, say, or hand over dangerous chemistry knowledge. The self-improvement question is whether a model can contribute to building AI itself, potentially speeding up its own successors. OpenAI says it's continuing to evaluate that capability as part of its ongoing safety work.
Worth keeping in mind: those High and Critical classifications come from OpenAI's own framework, making them a self-reported warning label rather than an outside audit.
TopicsOpenAI · GPT-6.2 Astra · GPT-6.2 Sol · GPT-6 Astra · ChatGPT Work · Codex
Source: cryptobriefing.com
Written by AI from verified reporting.
More stories
- Tech & AI techcrunch.com Caught by @putri.siregar49 0 viewsCalifornia DMV Grounds Waymo's Freeway Rides, and Waymo Pulls Them Everywhere
- Tech & AI nbcnews.com Caught by @taufik.utami96 1 viewApple Blames Internal Update for Outage Across iCloud, App Store and Store Systems
- Tech & AI livemint.com 1 viewOpenAI ships GPT-6.2 Astra after scrapping the previous model
- Tech & AI cryptobriefing.com 1 viewMicrosoft Rolls Out MAI Code 2 and Flash, Claims 50x Cost Edge

