Introducing GPT-6.1 Sol

OpenAI’s product post says GPT-6.1 Sol nearly matches GPT-6 Astra on agentic coding, computer use, and professional work at one-fifth standard input/output token prices; cached input is $0.10/M tokens. Cited evals include DeepSWE v1.1 matching Astra cheaper; GDP.pdf and AutomationBench gains versus Opus 5.5 and prior Sol; OSWorld 2.0 within 2.1 points of Astra at max effort; Terminal-Bench Science more than doubling prior Sol at under half cost, though Astra still leads science at 68.1%.

Factual errors at low effort fall from 11.4% to 7.7% on hard flagged chats. Alignment evals claim better transparency and constraint-following; no automated safety-reviewer bypass attempts. Available in ChatGPT Work and Codex for Plus–Edu; API gpt-6.1-sol at $2/$0.10/$10 per M input/cached/output; Ultrafast coming; not yet in Chat.

We’re reading the primary source as a cost-curve attack with a safety appendix.

If DeepSWE and AutomationBench hold outside the research environment caveats OpenAI itself footnotes, Sol is the model that lets agent fleets run without Astra invoices. That is useful AI for builders. The science bench honesty — Astra still wins — is welcome.

I’m watching the system card addendum and the Chat gap. Shipping Work/Codex first concentrates power users. Alignment charts on “challenging evaluations” need independent replication after a summer of reward hacking.

We’ll recommend Sol where price/performance matters and Astra where science hardness does — and we’ll keep the shelved 6.1 Astra story in the same frame.

OpenAI product announcement, September 29, 2026. Eval footnotes note research/API vs ChatGPT differences.