GPT-6.2 Astra and Sol Arrive in ChatGPT Work and Codex

OpenAI released GPT-6.2 Astra and Sol on November 17. A coding benchmark puts them half a point apart, but the cost per attempt tells a different story. Here’s how to choose between them.

GPT-6.2 Astra and Sol in ChatGPT Work and Codex, illustrated by a blue constellation and a golden sun.

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026, in ChatGPT Work, Codex and the OpenAI API. Astra is the model OpenAI recommends for demanding projects, such as turning a folder of reports into a presentation that holds up in a meeting. Sol has a claim on some demanding work of its own. In a test of changes to real codebases, it came within half a point of Astra at about one-fifth of the cost per attempt.

Where to find the new models

In ChatGPT, open Work and use the model picker. The regular Chat picker keeps its existing lineup. The rollout includes Plus, Pro, Business, Enterprise and Edu, with availability depending on the account and workspace settings. If the models haven’t appeared in the desktop app or Codex CLI, OpenAI recommends updating to the latest version.

Developers calling the API can use gpt-6.2-astra and gpt-6.2-sol. Signing in to Codex with a ChatGPT account draws on the plan’s allowance. Signing in with an API key uses API billing.

Where Astra earns a closer look

One of OpenAI’s launch examples is a valuation workbook. The earlier GPT-6 Astra put the revenue breakdown on one sheet and the calculation behind its share-price estimate farther down another. GPT-6.2 Astra placed those figures beside the table showing how the estimate changes when assumptions move. A reviewer can now trace the number they want to challenge without hunting through the file.

That example suggests a practical way to assess Astra. Give it last quarter’s presentation, this quarter’s figures and a short brief. Pick one chart in the finished deck. Can a colleague trace its headline back to the source figures and explain the change? That check captures something a polished slide or a correct total alone can miss.

Why a developer might start with Sol

Cognition’s FrontierCode 2.0 asks whether a code change meets an open-source maintainer’s criteria for merging it. Across its 132 tasks, a patch can lose all credit for missing a blocking requirement, even if it fixes the most obvious symptom. OpenAI reports a gap of just 0.5 points between Sol and Astra with both at maximum reasoning effort. Sol cost about one-fifth as much per attempt on that evaluation.

For a team choosing a model for routine maintenance, that makes Sol a sensible first trial. Take a recently fixed bug and give each model the original issue and the same starting commit. Run the regression tests, then compare how much editing the maintainer would need before merging each patch. A model that saves a few cents but adds half an hour of review is an expensive choice for that particular job.

What the API price difference buys

The following standard rates apply to short-context API requests at launch. Prices are in US dollars per million tokens.

Model Input Cached input Output
GPT-6.2 Astra $8.00 $0.40 $40.00
GPT-6.2 Sol $2.00 $0.10 $10.00

For a request with 10,000 uncached input tokens and 2,000 billable output tokens, the token charge comes to $0.16 with Astra and $0.04 with Sol, before any tool fees. That four-to-one ratio holds when token counts are identical. A coding agent can take more turns or generate more tokens on the same assignment, which is how a benchmark’s spending ratio can differ from the price table.

For teams already using GPT-6.1 Sol, the new Sol keeps the same standard token prices. Astra’s standard input and output rates fall 20% from GPT-6 Astra. The most revealing comparison for an upgrade is the cost of a result someone can actually use, including the time they spend checking it.

Sources