GPT-6.1 Sol, near Astra's level
On 29 September 2026 OpenAI released GPT-6.1 Sol, an update to GPT-6 Sol that it says nearly matches GPT-6 Astra on autonomous coding, computer use and professional tasks at one fifth of Astra's standard price. The API price is $2 per million input tokens, $0.10 cached and $10 output; the model is in ChatGPT Work and Codex for paid plans and in the API as gpt-6.1-sol.
Why it matters
It came the day after OpenAI said GPT-6.1 Astra would not be released, so the 6.1 step reached users only in the cheaper tier. In its comparisons OpenAI sets its mid-tier model beside its own flagship and Anthropic's Opus 5.5, not only beside its predecessor.
Prices. On OpenAI's page Astra costs $10 / $50 per million input and output tokens (cached input $1), GPT-6.1 Sol $2 / $10 (cached $0.10), GPT-6 Luna $0.10 / $0.50. The API changelog gives the same for prompts up to 272K input tokens, and lists a beta multi-agent mode in the same entry. The post promises Ultrafast (up to eight times faster token generation in Codex) 'in the coming days'. OpenAI's claims. DeepSWE v1.1: Astra's level at about one fifth of the cost and 6.4 points above GPT-6 Sol's best result. AutomationBench 1.0.6: 2.2 points above Opus 5.5 at medium effort at about a third of the cost. OSWorld 2.0 offline: 7 points above GPT-6 Sol at max effort, 2.1 below Astra. Terminal-Bench Science 0.1: Astra remains the highest (68.1%); GPT-6.1 Sol costs $5.47 per task at max effort against $23.21 for Opus 5.5 and $23.80 for Astra. The share of answers with a factual error on an internal test at low effort fell from 11.4% to 7.7%. On a test of broken search tools the model fails to report the problem in 2.1% of cases (GPT-6 Sol 4.9%, Astra 1.5%, Luna 28.7%); OpenAI recorded no attempts to bypass the automated safety review. A separate addendum to the GPT-6 Astra system card was published the same day. Availability. From 29 September in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu; not yet in Chat. What the record does not claim. All evaluations are OpenAI's. On the GDP.pdf test Opus 5.5 is taken 'with fallbacks' (Anthropic's routing to other models); on AutomationBench OpenAI itself writes that the cost shown for Fable 5.1 is understated because it leaves out the switches to a fallback model, which happened in about 40% of tasks. The post prints no date in its text: 29 September is from the API changelog.