OpenAI launches GPT-6.1 Sol at one-fifth Astra's token price

Why it matters

GPT-6.1 Sol makes OpenAI's cost-performance case directly to developers building agents: cached prompts are cheaper, while company-run tests claim near-Astra results on selected tasks. Actual savings will depend on workloads and token use.

OpenAI launched GPT-6.1 Sol on September 29th, positioning the upgraded model as a lower-cost option for coding and agentic work. CEO Sam Altman (@sama) highlighted its price in a post on X: standard input and output tokens cost one-fifth as much as GPT-6 Astra, while cached input tokens cost $0.10 per million. OpenAI's announcement says Sol nearly matches Astra on several coding, computer-use and professional-work evaluations.

The lower cached-input rate targets developers running agents that reuse long prompts. Cached input is material the model can reuse across requests instead of processing again at the standard input rate. OpenAI's $0.10 price is 95% below the $2 standard input rate and half the $0.20 cached-input price listed for GPT-6 Sol. The discount applies to cached tokens, not every token in a request; output remains separately priced at $10 per million tokens.

OpenAI lists GPT-6.1 Sol at $2 per million input tokens and $10 per million output tokens, against Astra's $10 and $50. Those rates make both categories one-fifth the price. They do not establish that every task will cost one-fifth as much: total spend also depends on token use, how much context is cached and the reasoning setting. OpenAI reports task-level comparisons too, and those figures vary by evaluation and setup.

On DeepSWE 1.1, a software-engineering test using real codebases, OpenAI says Sol matches Astra at roughly one-fifth of the cost and scores 6.4 percentage points above GPT-6 Sol's previous best result, with lower reasoning effort and cost. On GDP.pdf, which tests professional questions about complex documents, OpenAI says Sol scores above Opus 5.5 with fallbacks at less than half the cost per task, and approaches Astra at roughly one-fifth the cost.

On AutomationBench, which tests agents completing business workflows across 47 tools, Sol scores 2.2 percentage points above Opus 5.5 at medium reasoning effort, at about one-third the cost. On the offline set of OSWorld 2.0, Sol comes within 2.1 percentage points of Astra at maximum reasoning effort, at roughly one-seventh of Astra's task cost. OpenAI also reports that Sol more than doubles GPT-6 Sol's score on Terminal-Bench Science 0.1. At maximum effort, it says Sol costs an average of $5.47 per task, compared with $23.21 for Opus 5.5 and $23.80 for Astra; Astra still leads the test with a 68.1% score.

OpenAI says Sol reduced factual errors on its difficult factuality prompts from 11.4% for GPT-6 Sol to 7.7% at low reasoning effort. Across tested settings, its error rate stayed within 1.9 percentage points of Astra's at less than one-fifth the cost per task. The prompts came from de-identified conversations where users had already flagged an earlier model's error, which OpenAI says makes them unrepresentative of typical usage. The company also reports lower failure rates than GPT-6 Sol in several challenging alignment tests, including respecting explicit restrictions and avoiding unauthorized outcomes during agentic tasks. It observed no attempts to bypass an automated safety reviewer in its test.

These are OpenAI-reported evaluations, not independent tests of production results. The company says its GPT evaluations ran in its research environment or through its API, which can produce different outputs from ChatGPT because of differences in prompts, tools and reasoning effort. It says competitor scores came from publicly available reports. The results suggest where Sol may be useful; they do not settle how it will perform across a customer's own workflows.

OpenAI says GPT-6.1 Sol is available through its API under the model ID gpt-6.1-sol, and to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex. It is not yet available in Chat. The developer documentation lists a 1.05-million-token context window, a maximum output of 128,000 tokens, and support for tools including web search, file search, code interpreter and computer use. OpenAI says an Ultrafast version for Codex, with token generation up to eight times faster than standard speed, will follow.

The founder making the case has spent much of his career backing and building technology companies. Before OpenAI, Altman co-founded Loopt, a location-based mobile service that joined Y Combinator's first batch and was acquired by Green Dot in 2012; he later led Y Combinator. Sol's pitch is a developer-facing product decision about how much useful work customers can run for a given budget. The release gives builders a lower listed price and OpenAI's own benchmark evidence. Whether that changes deployment decisions will depend on results and bills outside the company's test suite.