OpenAI has released GPT-6.1 Sol, a model the company says is intended to give organizations a lower-cost option for agentic coding, computer-use tasks and multi-step business workflows. The model is available in ChatGPT Work and Codex for eligible paid users and as gpt-6.1-sol through the company’s API, creating another production choice for teams that need to balance model capability with recurring inference costs.
OpenAI positioned the model between GPT-6 Sol and its more capable GPT-6 Astra. It said GPT-6.1 Sol nearly matches Astra on agentic coding, computer use and professional work, while standard API prices are $2 per million input tokens and $10 per million output tokens. Cached input is priced at $0.10 per million tokens, a rate that matters for agents that repeatedly reuse a large instruction set or working context.
For enterprise development teams, the practical change is not simply a new model name. A lower cached-input price can make it more feasible to keep coding agents supplied with repository context, policies and task history across repeated calls. That can be relevant to workflows such as debugging, code review and change planning, although the company did not provide a total-cost estimate for a typical software-delivery deployment.
OpenAI reported that GPT-6.1 Sol improved over GPT-6 Sol on DeepSWE v1.1, a benchmark of software-engineering work in real codebases, by 6.4 percentage points at a lower reasoning setting and cost. It also said the model scored 2.2 percentage points above Anthropic’s Opus 5.5 on its AutomationBench test at medium reasoning effort, at roughly one-third of the cost. Those results are vendor-reported benchmark figures, rather than a guarantee of performance on an organization’s own code or processes.
The company also highlighted gains for computer-use work. On the offline set of OSWorld 2.0, OpenAI said the model exceeded GPT-6 Sol by seven percentage points at maximum reasoning effort while costing less than half as much per task. Computer-use systems can interact with applications through user interfaces, but organizations still need to determine which actions to allow, what confirmation steps are required and how activity will be audited.
GPT-6.1 Sol’s factuality results may be relevant to teams using the model with internal documents. OpenAI said the share of responses containing at least one factual error fell from 11.4% for GPT-6 Sol to 7.7% at low reasoning effort in a deliberately difficult set of de-identified conversations where users had flagged an earlier model error. The company cautioned that the test was not representative of typical usage.
OpenAI also said the model showed lower failure rates than GPT-6 Sol in evaluations covering transparency about broken search tools, compliance with explicit restrictions and avoidance of unauthorized outcomes during agentic tasks. It said it observed no attempts by the model to bypass an automated safety reviewer in those tests. The company’s system-card addendum provides the underlying evaluation context, but enterprises will still need to apply their own access controls and test agent behavior before connecting production systems.
GPT-6.1 Sol is available now to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex, according to OpenAI; it is not yet available in Chat. The company also plans an Ultrafast version for Codex in the coming days, with up to eight times faster token generation than its standard speed.
For teams building agents that must reason over repeatable context, the new price structure is the immediate operational point. Lower token costs can expand the number of workflows worth piloting, but production value will still depend on reliable integrations, permissions, evaluation and human oversight rather than benchmark performance alone.