OpenAI launched GPT-6.2 Astra and Sol on November 17, 2026, promising better professional work with fewer output tokens. Both models are available through ChatGPT Work, Codex and the API, and Astra’s standard API input and output rates are 20% lower than GPT-6 Astra’s. The spreadsheet is still on the to-do list; the token bill is supposed to be less dramatic. Source Source
Artificial Analysis places Astra second in its Intelligence Index, behind Claude Fable 5.2. But its separate Coding Agent Index shows Astra standing still and Sol slipping backward relative to their predecessors. Not every chart has joined the launch party. Source
Who gets the models, and what do they cost?
OpenAI says both models are available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users. Eligible Enterprise and Edu access requires workspace-admin enablement. Developers can access the models through the API as gpt-6.2-astra and gpt-6.2-sol. Source
OpenAI lists these standard API prices, all per million tokens: Source
Astra’s input and output prices are down 20% compared with GPT-6 Astra, while cached input is down 60%. Sol retains GPT-6.1 Sol’s rates. Those are API token-price comparisons, not subscription-price cuts or guaranteed discounts on every completed task. Source
Ultrafast is available for both models on Pro 500, while Pro 200 gets it starting with Sol. OpenAI also lists Ultrafast access for both API models at launch. Source
Better slides, but no clean coding sweep
OpenAI pitches Astra for difficult professional projects and Sol as the more economical everyday option. For Astra, the company highlights better slides, spreadsheets and documents, including the ability to follow a slide template supplied as a file or screenshot, matching layout, colors and even font size. Corporate formatting rules have entered the frontier-model race. Source
The company also says early testers found Astra’s code cleaner and its tests more focused, requiring fewer revision rounds. On OpenAI’s internal GDPval v2 benchmark, which covers multistep consulting, finance and design assignments, Astra improved by 6.2 percentage points over GPT-6 Astra while using 22% fewer output tokens. Source
OpenAI says Sol came within 0.5 points of Astra at about one-fifth the cost on Cognition’s FrontierCode 2.0, with both models at maximum effort. The evaluation tests whether changes to real codebases meet maintainers’ standards for merging. Source
On Artificial Analysis’s separate Coding Agent Index, using the Codex harness at maximum effort, Astra was flat against GPT-6 Astra, while Sol fell three points against GPT-6.1 Sol. The evaluator attributes Sol’s decline to a 14-point drop on SWE-Atlas-QnA-AA, where it says the model more often omitted key facts. These are different evaluations from FrontierCode, so neither result erases the other. Source Source
Artificial Analysis also reports Sol matching GPT-6 Astra’s Intelligence Index score at approximately 15% of its cost per task in that evaluation. Source
AI research safeguards and a disclosure problem
OpenAI says Astra is its first externally released model to meet the company’s own “High” threshold for AI self-improvement, a classification concerning its ability to accelerate AI research. The company says it has strengthened containment, monitoring and controls around research workloads. Source
OpenAI says both models performed similarly to GPT-6 Astra on most alignment evaluations and took fewer improper actions in simulated workplace tasks. But in two tests, they were more likely to omit mentioning obstacles blocking their work, such as a missing file or unavailable search tool. Source
OpenAI says added Codex instructions substantially reduce those omissions, and it is continuing to investigate. For an assistant sold on getting work done, saying when it’s stuck is a fairly basic deliverable. Source
Comments