On 29 September OpenAI released GPT-6.1 Sol, an upgrade to GPT-6 Sol. Per the company it nearly matches GPT-6 Astra on agentic coding, computer use and professional tasks at a fifth of Astra's token prices. It is available in ChatGPT Work, Codex and the API, but not yet in regular chat.
- API pricing: 2 dollars per million input tokens, 0.10 cached, 10 output.
- At low effort the share of answers with a factual error falls from 11.4 to 7.7 percent, on deliberately hard conversations.
- An Ultrafast variant with up to 8 times faster generation in Codex is announced for the coming days.
One week. That is how long GPT-6 Sol lasted as the latest version before 6.1 arrived.
The pace is news in itself, but not the main news. More important is how OpenAI arranges its own family: Astra is the top, Sol is the workhorse, and this Sol upgrade is measured by how close it gets to Astra and how much cheaper it is.
The number I like most is the modest one: factual errors, from 11.4 to 7.7 percent. OpenAI itself points out that the test is on conversations where people had already caught an older model making a mistake. So a deliberately hard case, not the usual one. And still, almost one answer in thirteen there contains at least one error.
Something else shows in their tables: Anthropic's Opus 5.5 is now the yardstick they measure against. On AutomationBench GPT-6.1 Sol is 2.2 points above it at medium effort and at roughly a third of the cost, per OpenAI.
For the hardest scientific work OpenAI itself points you to Astra, which holds 68.1 percent on Terminal-Bench Science. A maker rarely tells you when to pay more. Here it does.