OpenAI introduces GPT-6.2 Astra and Sol: available in ChatGPT Work, Codex and the API
OpenAI introduced GPT-6.2 Astra for its hardest projects and GPT-6.2 Sol for everyday work. The company says both models can be used in ChatGPT Work and Codex on the Plus, Pro, Enterprise and Edu plans, and by developers through the API. The comparison results were published by OpenAI itself.
OpenAI released two new models: one for hard work, one for everyday tasks
OpenAI presented GPT-6.2 Astra as its new frontier model for the hardest projects and GPT-6.2 Sol as an efficient, versatile model for everyday work. The company says both reach better results on demanding tasks while producing fewer output tokens, which means more finished work inside your usage limits.
The announcement describes access this way: from launch day, every user on the Plus, Pro, Enterprise and Edu plans can pick the models inside ChatGPT Work and Codex. For eligible Enterprise and Edu users, access follows once a workspace admin turns it on. Developers can call gpt-6.2-astra and gpt-6.2-sol through the API.
Astra focuses on document and interface quality, Sol on efficiency
OpenAI says Astra's slides, spreadsheets and documents sound less machine-made and stay closer to the templates you supply and the habits of fields like law, sales, finance and accounting. The same text says Astra writes cleaner code and builds front ends that work better, especially for websites, games and 3D design.
The comparisons rest on OpenAI's own statements. On Agents' Last Exam, a test of multistep professional work, the company reports that Astra scores 5.2 points above GPT-6 Astra at under a third of the cost and leads Claude Fable 5.2 by 5.5 points. OpenAI also notes that the Fable 5.2 results leave out nine tasks that used a fallback model. On Harvey LAB-AA, which covers legal work, Astra is said to more than double Fable 5.2's pass rate.
For Sol the claim is that it lands within half a point of Astra on Cognition's FrontierCode 2.0 at roughly a fifth of the cost. On LifeVal, a test of everyday tasks, it is said to beat GPT-6 Astra at around a seventh of the cost. OpenAI adds that GPT-6.2 Sol lets users finish more tasks on the same paid plan than GPT-6.1 Sol did.
On pricing, the company says Astra's API token rates fall compared with GPT-6 Astra, with a deeper cut for cached input, while Sol keeps GPT-6.1 Sol's rates. To get more cache reuse, OpenAI suggests leaving prompt prefixes unchanged and appending new instructions at the end. Pro 500 subscribers can use Ultrafast speed with either model, and Ultrafast reaches Pro 200 subscribers for the first time, starting with Sol.
Choosing a model is now about cost and efficiency as much as quality
The weight of the announcement is not only a stronger model: OpenAI puts the same work done with fewer tokens up front. In agent-based workflows a task means many steps and tool calls, so token efficiency shows up directly in a company's monthly bill. That is our reading.
The two-tier setup is worth noting as well. Astra is positioned for complex, long jobs and Sol for frequent ones, and OpenAI says many of its own staff use Sol as their default. That split could let a single workflow give the hard steps to the stronger model and the routine steps to the cheaper one.
There are cautions. Every benchmark result we cite comes from OpenAI's announcement, and we have not checked the comparisons with a rival's model against an independent source. OpenAI also says both models ship with tighter safeguards around cybersecurity and biological topics, and that the cybersecurity safeguards also restrict some legitimate vulnerability research and security-testing workflows. Finally, the company reports that in two tests the models were more likely to leave out an obstacle blocking their work, such as a missing file; it says it added instructions to Codex that cut this down and is still investigating.
What it means for businesses in Türkiye
This section is our reading, not a fact from the source. The announcement describes access by plan type (Plus, Pro, Enterprise, Edu) and through the API. We cannot tell from it when the models appear on accounts in Türkiye, so the safest step is to check the model list in your own account.
For a business that already uses ChatGPT or the API, the practical move is a small trial: run a report, quote or spreadsheet job you do often through both models and compare quality and the share of your usage it consumes. OpenAI's figures do not mean you will see the same result on your own work, and a measurement on your own data is worth more for the decision. That too is our suggestion.
The UNALSOFT view
New model releases bring back the question of which step in an agent-based project should get which model. We read this news as one example of why the cost and quality balance should be measured at the start of workflow design. In our agentic AI work we settle model choice through small tests on your own data. The announcement gives no Türkiye timeline, so this is a development to plan around, not an urgent decision.
Sources
Which model should run which step in your workflows?
A short conversation is enough to define your processes and the cost-quality balance together.