AI

GPT-6.2 Astra and Sol reach ChatGPT, and the cheaper one nearly ties on code

Susan Hill
Add us on Google

OpenAI has upgraded both of its main models for ChatGPT and Codex, Astra and Sol, and the most interesting thing about the new versions is how small the gap between them has become. Astra is the flagship, built for the hardest projects. Sol is the lighter, cheaper model for everyday writing and coding, and on one demanding coding test OpenAI says it landed within half a point of Astra.

For anyone paying for ChatGPT, the change shows up in the usage meter. OpenAI says both models finish demanding tasks with fewer output tokens, the units of text a model produces as it works, so the same plan stretches further before you hit its limit. Developers also get a price cut on the flagship.

GPT-6.2 Astra’s party trick is visual imitation. Give it a screenshot of a slide and ask for the rest of the deck in the same style, and OpenAI says it will match the layout, the colors and even the font size. The company says its slides, spreadsheets and documents “read less like AI” and follow the conventions of the field they are written for. In Codex, OpenAI’s coding tool, Astra is meant to be better at changing and testing an existing project than at starting from a blank page. “Early testers reported cleaner code style and more focused tests, with fewer rounds of revision,” the company said.

The Sol result is the one to watch. On FrontierCode 2.0, a benchmark from Cognition that checks whether changes to real codebases are good enough for their maintainers to merge, Sol came within half a point of Astra at about one-fifth of the cost, with both models at maximum reasoning effort. For plenty of everyday coding, the cheaper model may be all you need.

In the API, Astra costs $8 per million input tokens and $40 per million output tokens, 20% less than GPT-6 Astra and the same price as Anthropic’s Claude Fable 5.2. Cached input, text the model has already seen in a session, drops to $0.40 per million from $1. That matters for long agent sessions that reread the same material again and again. Sol keeps its price of $2 and $10.

The release is not a clean win. Early reviews are mixed: developer Simon Willison noted that Astra doesn’t score quite as high as Claude Fable 5.2 on several benchmarks, and some testers find the way it communicates confusing. Theo Browne, who tried it before launch, described a long task that stalled two hours in because the model stopped to ask him about an internal file name he could not possibly know. OpenAI’s figures come from its own tests and partner benchmarks, and the half-point result for Sol is a single coding test, not a verdict on everything.

Both models are available now in ChatGPT Work and Codex on the Plus, Pro, Enterprise and Edu plans, and through the API as gpt-6.2-astra and gpt-6.2-sol. OpenAI’s announcement does not mention the free tier. The $200-a-month Pro plan gets OpenAI’s Ultrafast mode for the first time, on Sol only, while the $500-a-month Pro plan gets it on both models. OpenAI says Ultrafast generates text up to 4.5 times faster than the standard rate, but it uses more of your allowance. Access follows ChatGPT’s paid plans, which are not sold in every country.

The launch lands a week after Anthropic’s Claude Fable 5.2 and two months after GPT-6 Astra arrived in September. Next up are Enterprise and Edu workspaces, which get Ultrafast on both models once an admin switches it on.

Tags: , , , , ,

Add us on Google

Discussion

There are 0 comments.