OpenAI releases top-end Astra and everyday Sol together
OpenAI on Nov. 17 (Nov. 18 Japan time) released two new AI models, "GPT-6.2 Astra" and "GPT-6.2 Sol." It positioned Astra as a new frontier model for the hardest projects and Sol as an efficient, versatile model for everyday work. ChatGPT Plus, Pro, Enterprise and Edu users can use them in "ChatGPT Work" and the AI coding tool "Codex," and they are available to developers through the API from the same day. Astra's API price is $8 for input and $40 for output per million tokens, 20% lower than the previous top model, "GPT-6 Astra." Sol costs $2 for input and $10 for output, carrying over the prices of GPT-6.1 Sol.
The company announced the release on its official X account at 2:54 a.m. Nov. 18 Japan time. It said both models produce better results on demanding tasks while using fewer output tokens (the units in which AI processes text), letting users get more done within their usage limits.
Some of you have been asking very politely. Some of you have just been asking. GPT-6.2 Astra and Sol are now available. Super efficient and built to perform where it counts—the real world.
Regarding 6.2, on Nov. 1 the label "gpt-6.2-astra" appeared in Codex usage records, and on Nov. 4 The Information reported that a new Astra and Sol could be released soon. The same day, a leak said the two names had appeared on Google Cloud. The release formally announces models that had been the subject of leaks and reports.
OpenAI says the models undercut earlier models and rivals on cost per task
OpenAI presented its own evaluations comparing performance and cost. On "Agents' Last Exam," which tests multistep professional work, Astra scored 5.2 points higher than GPT-6 Astra at less than a third of the cost, the company said. Against Anthropic's Claude Fable 5.2, it said Astra scored 5.5 points higher at 5% of the cost. The page's text gives only the differences, not the scores themselves. The company noted that on this benchmark's leaderboard, Fable 5.2's results exclude nine tasks that used a fallback model.
In science, OpenAI said Astra set a new high on "Terminal-Bench-Science 0.1," which tests how well models carry out research work in a terminal, scoring 9 percentage points higher than GPT-6 Astra. The cost was less than a third, it said. According to a footnote on the page, Fable 5.2's result is as measured by Artificial Analysis with the "mini-swe-agent" harness, and Anthropic reported a score of 74.6% at max reasoning in Claude Code. OpenAI said Astra also set a new high on the "Vals Science Index," which asks models to interpret biological data.
For Sol, the company said that on Cognition's "FrontierCode 2.0," it came within 0.5 points of Astra at about one-fifth the cost (both at max reasoning). On LifeVal, which the company recently released to measure how well models handle everyday errands, Sol outscored GPT-6 Astra at about one-seventh the cost, it said. On the internal "GDPval v2" evaluation, Astra rose 6.2 points from GPT-6 Astra while using 22% fewer output tokens. Sol rose 5.3 points from GPT-6.1 Sol.
All of these are the company's own evaluations, and no third-party verification has been shown.
OpenAI highlights gains in documents, legal work and 3D
OpenAI said it improved the quality of slides, spreadsheets and documents with Astra. It said Astra follows the templates it is given and the conventions of fields such as law, sales, finance and accounting, with fewer generic takeaways and less filler. The company said Astra also writes cleaner code and builds better websites, games and 3D experiences.
On "Harvey LAB-AA v1.1," which uses legal assignments, OpenAI said Astra more than doubled Claude Fable 5.2's pass rate at about one-sixth the cost. Astra for Law, its configuration for lawyers, scored even higher, the company said. On Artificial Analysis's "AA-AnalystAgent," which tests finding and calculating figures from spreadsheets and reports, Astra set a new high, answering 62.5% of questions correctly on all five independent attempts, it said. On "AA-3D," where people compare two 3D creations and vote, Astra ranked first with 1,318 (a head-to-head Elo score), followed by Claude Fable 5.2 at 1,271 and Sol at 1,259, the company said.
The official page also carries customer comments. An engineer at Basis, which automates accounting work, said the average time for long-horizon accounting work fell by 50%, and by up to 73% for the most complex work. Engineers at Ramp and Jump Trading also praised Astra for system design judgment and optimization work.
Cyber restrictions tightened; 6.2 still to come to Daybreak
OpenAI said both models make a significant jump in cybersecurity capabilities. They find and exploit vulnerabilities more effectively, write more secure code and produce better patches, it said. Under its safety framework, the "Preparedness Framework," the company assessed that the models meet the highest "Critical" threshold for cybersecurity, as GPT-6 Astra does.
As a result, it launched them with tighter safeguards for higher-risk cybersecurity assistance. The company acknowledged that while these safeguards reduce misuse risk, they also restrict some legitimate vulnerability research and security testing work. Through "Daybreak," its program for defenders, it currently offers GPT-6 Astra and GPT-6 Cyber, and said it will work to bring the 6.2 models there. The cybersecurity results it published were measured in internal testing configurations with safeguards and classifiers disabled, it said.
The company said the models meet the "High" threshold for biological and chemical capabilities. It said Astra is its first externally released model to meet the "High" threshold for "AI Self-improvement," which measures the ability to accelerate AI research. OpenAI said it strengthened monitoring, including "Control Room," a system that tracks research workloads and can halt further model output when needed.
In evaluations of whether AI acts in line with human intent (alignment), both models performed similarly to GPT-6 Astra on most items and took fewer improper actions in simulated workplace tasks, the company said. However, in two tests they were more likely to omit telling users about obstacles that blocked their work, such as a missing file or an unavailable search tool. The company said it added instructions in Codex that substantially reduce these omissions and is continuing to investigate the behavior. It said detailed results are in the system card.
Available on four plans including Plus; API price for Astra cut 20%
In ChatGPT, all Plus, Pro, Enterprise and Edu users can use both models in ChatGPT Work and Codex from the same day. Enterprise and Edu users get access once their workspace admin enables it. The official page does not mention the Business, Free or Go plans.
Pro 500 subscribers can use the "Ultrafast" speed setting with either model. The company is bringing Ultrafast to Pro 200 subscribers for the first time, starting with Sol. In the API, the model names are gpt-6.2-astra and gpt-6.2-sol, and Ultrafast is available for both models at launch.
API prices per million tokens are $8 for input, $40 for output and $0.40 for cached input (the price when the same input is reused) for Astra. That is 20% lower than GPT-6 Astra's $10 for input and $50 for output, with a 60% discount on cached input. Sol is $2 for input, $10 for output and $0.10 for cached input, carrying over GPT-6.1 Sol's prices. The company said that, combined with better token efficiency, Astra costs significantly less per task than GPT-6 Astra. To make use of caching, it recommends keeping prompt prefixes unchanged and appending new instructions. The official page does not state the context window.
- GPT-6 Astra
- GPT-6.2 Astra
- GPT-6.2 Sol






