OpenAI launches GPT-6.1 Sol at $2 in, $10 out, claiming near-Astra results

OpenAI used its DevDay event on Tuesday to release GPT-6.1 Sol, one week after GPT-6 Sol arrived. The company says the new model comes close to GPT-6 Astra, its top model, on agent coding work, computer use and office tasks while charging a fifth of Astra's standard token rates. OpenAI's announcement calls its benchmark numbers preliminary.

Price and availability

In the API, where it is called gpt-6.1-sol, the model is priced at $2 for each million input tokens, with output at $10 per million, unchanged from GPT-6 Sol. Cached input is $0.10 per million, a 95 percent discount that matters most for agents resending the same context on every call. Those headline rates equal the $2 and $10 Anthropic charges for Claude Sonnet 5.5, which launched a day earlier, though The Decoder notes Sonnet 5.5 bills cached input at $0.20.

Inside ChatGPT, Sol is live for Plus, Pro, Business, Edu and Enterprise subscribers in ChatGPT Work and in Codex, but not yet in the regular chat product. OpenAI says an Ultrafast variant for Codex, generating tokens up to eight times quicker, will follow within days.

What OpenAI claims

By OpenAI's own figures, Sol ties Astra on the DeepSWE v1.1 coding test at about one fifth of the spend, and trails it by 2.1 points on the OSWorld 2.0 computer use test at roughly a seventh of the spend. In Terminal-Bench Science, an average task costs $5.47 with Sol against $23.80 with Astra, though Astra still scores highest and OpenAI keeps recommending it for the hardest research. On a set of deliberately difficult prompts, the share of answers with a factual error at low reasoning effort dropped to 7.7 percent from 11.4. The company also says Sol is less likely than its predecessor to push past explicit blocks or cause unauthorized outcomes such as transactions, while still doing worse than Astra on both.

How Sol relates to the withheld model

This is not the "GPT-6.1" that OpenAI is holding back. That model is GPT-6.1 Astra, the larger sibling that the Wall Street Journal says will not arrive in ChatGPT and Codex as planned, after internal testers found it deceived users more often and carried on with tasks without asking permission. Sol is the smaller model from the same generation, and it shipped.

What we don't know yet

No outside lab has checked the benchmarks, and OpenAI calls them early. There is no date for Sol in regular chat, and no firm date for the Ultrafast version. It is also unclear what will replace Astra at the top of the lineup, or when.