Illustrative image: LNA
SAN FRANCISCO, November 17, 2026 — OpenAI has released GPT-6.2 Astra and GPT-6.2 Sol, a pair of upgraded models it says do more professional work for less money, and it has flagged Astra as the first model it has released publicly that crosses its own “High” risk threshold for speeding up AI research.
Astra is the new frontier model for hard, multi-step jobs such as research, coding and polished slides, spreadsheets and documents. Sol is the faster, cheaper model for everyday work. Both went live on Tuesday for ChatGPT Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and for developers through the API, OpenAI said in its announcement.
Price is the headline. In the API, GPT-6.2 Astra costs US$8 per million input tokens, US$0.40 for cached input and US$40 for output, which OpenAI said is 20 per cent cheaper on input and output than GPT-6 Astra, with a 60 per cent cut on cached input. GPT-6.2 Sol keeps its predecessor’s rates of US$2, US$0.10 and US$10.
OpenAI’s own benchmark claims are bold. It says Astra scores 5.2 points higher than GPT-6 Astra on Agents’ Last Exam, a test of long professional workflows, at less than a third of the cost, and 5.5 points ahead of Anthropic’s Claude Fable 5.2 at about 5 per cent of the cost. It also calls Astra “the best model in the world for scientists”.
Independent testers were more mixed, according to The Decoder. Artificial Analysis placed Astra second on its Intelligence Index, just behind Claude Fable 5.2, while leading on cost for performance. But it found no gain over GPT-6 Astra in its coding-agent index, and a high rate of guessing on very hard factual questions. Epoch AI rated Astra level with Fable 5.2 on its capabilities index, though weaker on real research tasks. The Verge also reported the launch.
The safety notes deserve attention. OpenAI said both models meet the “Critical” threshold for cybersecurity under its Preparedness Framework and the “High” threshold for biological and chemical capabilities, and that it is deploying them with its “strongest safeguards to date”. Astra is also its first externally released model to reach “High” for AI self-improvement, which OpenAI likened to giving every one of its researchers an extra, highly capable research engineer. It said it has added more oversight and monitoring in response.
Why it matters: the frontier race is now as much about cost per task as raw capability. Cheaper top-tier models make it easier for businesses, including Malaysian start-ups and companies paying in US dollars, to run AI agents all day. The self-improvement flag is a reminder that the labs themselves now treat faster AI research as a risk to be managed.
What to watch next: whether independent coding results catch up with OpenAI’s claims, how Anthropic and Google respond on price, and how the tighter cyber safeguards affect legitimate security researchers.
Sources: OpenAI; The Decoder; The Verge.

