Skip to content

OpenAI Launches GPT-6.2 Astra and Sol for ChatGPT Work, Codex and the API

News
Cobalt-blue desktop workstation and silver laptop side by side under amber and cyan light.
OpenAI released GPT-6.2 Astra and Sol on November 17. Both are now in ChatGPT Work, Codex and the API. Astra is built for hard jobs. Sol costs less and is meant for day-to-day work.
Both models are open to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex. Pro 500 users get the faster Ultrafast mode for both. Pro 200 users get it for Sol. Enterprise and Edu users can get the same access if their admin turns it on. The API also offers Ultrafast for both models. OpenAI lists access by plan.

How GPT-6.2 Astra and Sol differ

Astra is meant for tasks with more steps and more to check. It can write code or make slides, sheets and reports from a given template. To show what changed, OpenAI gave the old and new Astra the same set of books for a made-up firm. Both had to build a sheet that tracked sales by segment. The sheet also had to check that the firm’s three main statements matched.
On the Agents' Last Exam test, the new Astra beat the old one by 5.2 points. OpenAI says it cost less than a third as much per task on that test. Your cost can change with the size of the prompt and the work the model has to do.
Sol is built for the tasks that come up each day. It can work with text, pictures and tools in one job. OpenAI shared a code test from Cognition where Sol came within half a point of Astra. It cost about a fifth as much on that test with both models set to their highest reasoning level. OpenAI also put out the lower-cost GPT 6.1 Terra in October.

What the two models cost

GPT-6.2 Astra costs $8 per million input tokens, $0.40 for cached input and $40 for output. Its input and output rates are down 20 percent from the old Astra. Sol keeps the prior rates of $2, $0.10 and $10 per million tokens.
Apps can call the models through Responses or Chat Completions. Tools need the Responses API. The guide for developers notes that edits to the start of a chat can make a model lose its earlier reasoning. The model may then need to do that work again.
OpenAI also began a test of two-hour and eight-hour windows for API jobs that can wait. Both come with a lower rate. The API change log lists the rates and models that take part.

What OpenAI found in tests

GPT-6.2 Astra is the first model OpenAI has released that reached its High risk level for the ability to speed up AI research. OpenAI says it now tracks more of that work and can pause some jobs. In other tests, both models at times failed to say when they got stuck. A file could be gone or a search tool could be out of reach. OpenAI added rules in Codex that cut down on those misses in its tests. The system card has more on the findings.
loading

Loading