OpenAI introduced GPT-6.1 Sol at DevDay on Tuesday. The interesting part isn’t in the benchmark bars. It’s in the price table and in the choice of models it gets compared against.
The prices sit on the same line
GPT-6.1 Sol costs $2 per million input tokens and $10 per million output tokens through the API. That is, to the cent, what Claude Sonnet 5.5 costs, and what GPT-6 Sol cost before it. On cached input OpenAI goes lower: 10 cents per million against Sonnet 5.5’s 20, and half what its own predecessor charged.
The model is available now to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and in Codex, and through the API as gpt-6.1-sol. Not yet in ordinary Chat. A variant called GPT-6.1 Sol Ultrafast is due in the coming days, with up to eight times the token generation speed in Codex.
The pitch is proximity to the big model: OpenAI says Sol 6.1 nearly matches GPT-6 Astra on agentic coding, computer use and professional work, at a fifth of Astra’s standard prices.
The comparisons are to Opus 5.5 and Fable 5.1
Who OpenAI puts in its own charts is worth noticing. On DeepSWE v1.1, which tests real software work in real codebases, Sol 6.1 draws level with Astra and lands 6.4 percentage points above GPT-6 Sol at a lower reasoning effort. On GDP.pdf, where models answer professional questions from complicated PDFs, it scores higher than Opus 5.5 with fallbacks at less than half the cost per task. On AutomationBench, 47 tools across six business functions, it sits 2.2 points above Opus 5.5 at medium effort and costs about a third as much.
A footnote under that same chart deserves attention. The datapoint for Claude Fable 5.1 understates its real cost, OpenAI writes, because it leaves out the cost of fallbacks. Those happened on roughly 40 percent of tasks. So OpenAI is flattering a competitor’s number relative to its own measurement, and says so.
On computer use (OSWorld 2.0, offline set) Sol 6.1 lands within 2.1 points of Astra at maximum effort, for about a seventh of the cost per task. On Terminal-Bench Science 0.1 it more than doubles GPT-6 Sol’s score at maximum effort, averaging $5.47 per task.
Every competitor figure comes from publicly available reports, OpenAI says, while its own come from an internal research environment or the API. That caveat is standard, and it applies here as everywhere.
Price has stopped being the argument
Two dollars in, ten out: until Tuesday that line was the Sonnet line. Now it’s also OpenAI’s workhorse, and the cache is cheaper. If you run a pipeline that reuses a lot of context, those ten cents show up on the invoice immediately.
The second observation matters more. OpenAI no longer compares its mid-tier model to Sonnet. It compares it to Opus 5.5 and Fable 5.1, which is what Anthropic sells for long agentic runs and knowledge work. That’s positioning, not measurement: the claim is that you don’t need Anthropic’s big model because OpenAI’s middle one will do. Whether it holds gets settled on day three inside a real repository, not on a bar chart.