OpenAI introduced GPT‑6.2 Astra and GPT‑6.2 Sol on November 17. Both models are coming to ChatGPT Work, Codex, and the API. Sol is the default for new Work sessions on supported personal plans. Astra is available for projects that involve bringing a lot of material together and checking how the pieces fit.
For developers, one of the main differences is price. For the same number of tokens at standard API rates, Sol costs one quarter as much as Astra.
Where to find GPT‑6.2 in ChatGPT
Open Work in ChatGPT to find GPT‑6.2 Astra and Sol. Work is available on the web and in the desktop and mobile apps. Regular Chat has its own model list, so switch to Work if you have been looking for the new Astra and Sol there.
ChatGPT Learn lists Plus, Pro, Business, Enterprise, and Edu. An administrator needs to enable the models for Enterprise and Edu workspaces. If the options are missing, check your mode, plan, and workspace settings. Accounts may get access at different times while the rollout is in progress.
Look for the model control underneath the message box. If it shows a Power slider, open Advanced to choose a specific model and reasoning effort. The Codex app also has a model picker. In the terminal version, /model opens it. Work and Codex share subscription usage.
When Sol is enough
Sol is a good fit when you already know what the finished result should look like. You might have notes from a meeting and want a short account of the decisions and open questions. Or you might have a complete draft that needs to read more easily while keeping every figure and name intact. The default model gives you a straightforward place to start on those jobs.
On supported personal plans, a new Work chat starts with Sol at Light reasoning effort. Enterprise workspaces with access enabled start at Medium. Reasoning effort controls how much work the model can do before producing its response. Higher effort may add time to a simple request. When several sources disagree, that extra room lets the model consider more of the details.
A clearer request can make a bigger difference than changing models. For meeting notes, ask for decisions in one list and questions in another. Tell the model to include owners and deadlines wherever the notes name them. That gives the result a useful shape and makes missing details easier to spot.
When to bring in Astra
When a presentation draws on several reports and spreadsheets, much of the work happens before the first slide. The reports might use different names for the same metric, or a month might be missing from a spreadsheet. Someone needs to reconcile the sources and notice where they disagree. OpenAI positions Astra for projects like this.

For that presentation, ask for the filename and page or sheet name alongside each figure. Whoever receives the deck can then trace the numbers back to the material. A well-designed chart is easy to read, but its appearance tells you little about whether the values came from the right period.
The distinction also applies to coding. Sol can handle a clear edit in an existing codebase. Astra is an option when a problem could start in several parts of an application and you need to trace how those parts work together. Judge the result by the finished change and the tests that show how it behaves.
What Astra and Sol cost in the API
API usage is measured in tokens. These are chunks of text and other data the system counts while processing a request. There are separate rates for fresh input, input reused from a cache, and output. The rates below apply to API calls. Usage through ChatGPT follows the terms of your subscription plan.
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT‑6.2 Astra | $8 | $0.40 | $40 |
| GPT‑6.2 Sol | $2 | $0.10 | $10 |
The table shows rates per million tokens at Standard speed for requests of up to 272,000 input tokens. A request with 10,000 input tokens and 2,000 billed output tokens comes to $0.16 on Astra or $0.04 on Sol, assuming uncached input and no additional services.
A short response on the screen can cost more than its length suggests. The tokens a model spends on reasoning are billed as output too. If an application retries a request or adds search results, the overall usage can change again. The usage data returned by the API is a better basis for estimating what a job costs.
Both models also offer Ultrafast mode in the API for faster text generation. Its rates at launch are three times the Standard rates. Ultrafast availability in Work depends on your plan and selected model.
How large requests are billed
The documentation lists a 1,050,000-token context window for both models. Context is the room available for the material a model receives and the response it generates. The API limits are up to 922,000 tokens of input and 128,000 of output. Codex can use a different default context size, so those API limits and the initial Codex setting describe different things.
Once a request exceeds 272,000 input tokens, higher rates apply to the entire request. Input and cache rates double, and output costs one and a half times the short-context rate. A large collection is worthwhile when the model needs to use all of it. If your question is about one section of a document, send the section and the material it depends on.
What changes if you already use Sol
GPT‑6.2 Sol keeps GPT‑6.1 Sol’s standard rates for input, cached input, and output. The new model IDs are gpt-6.2-sol and gpt-6.2-astra. Both work with the Responses API and Chat Completions. The documentation directs developers to the Responses API for tool calls with these models.
If you use Sol for a lot of similar requests, compare the two versions with examples from your own workload. For a job that extracts fields from documents, look at the cost alongside the number of missing or incorrect fields. A model may generate fewer tokens, but the savings disappear if someone has to fix every other record by hand.

Join the conversation
Stay on topic and respect other readers. Your first comment may appear after editorial review.