OpenAI releases GPT-6 Sol and Luna with lower API prices and fewer errors

OpenAI released GPT-6 Sol and GPT-6 Luna on 22 September 2026, expanding its model range with lower API prices and reported gains in professional work and factuality, according to its announcement of the models. The launch follows GPT-6 Astra, which OpenAI introduced earlier in September as the more capable option in the family.

OpenAI’s new Sol and Luna API rates

OpenAI lists Sol at $2 per million input tokens and $10 per million output tokens, down from $4 and $20 for GPT-5.6 Sol. Luna costs $0.10 for input and $0.50 for output, compared with $0.20 and $1.20 for GPT-5.6 Luna. These are separate charges for the material sent to a model and the material it generates. OpenAI describes the earlier GPT-5.6 rates as promotional pricing; an OpenAI spokesperson told VentureBeat that the new Sol and Luna rates are permanent, rather than introductory offers.

For Luna, the input price is 50% lower and the output price 58.3% lower than the listed GPT-5.6 rates, VentureBeat reported. The publication also put Sol alongside Anthropic’s Claude Sonnet 5 at the same $2 input and $10 output rates. Anthropic released Claude Opus 5.5 on 22 September at $4 for input and $20 for output per million tokens, according to VentureBeat. Those token prices matter to developers choosing which model to call repeatedly, although the amount of text a task consumes also affects its bill.

OpenAI attributes its lower prices to improvements in inference, the process of generating an answer, and caching, which keeps previously processed material available for reuse. Caching is akin to keeping a worked-out page at hand rather than doing the same work again when that page is needed. OpenAI says the changes help it serve Sol and Luna more cheaply, but its published performance results also depend on the model and effort setting used for each task.

Summarising a document could cost less through Luna at the stated API rates than through its predecessor, if the same amounts of text were sent and generated. A task that repeatedly uses earlier material could also benefit from the caching improvements OpenAI describes. Neither possibility removes the need to decide whether the cheaper model gives an adequate answer.

Sol scores 33.2% on AutomationBench

OpenAI tested agents on AutomationBench 1.0.6, which uses end-to-end workflows involving 47 tools across functions including sales, support, finance and HR. In the company’s results, GPT-6 Sol at xhigh effort scored 33.2% and cost $0.27 per task. Claude Opus 5 at max effort scored 26.9%, with a cost per task that OpenAI put at 11.1 times Sol’s. The cost comparison concerns runs of that benchmark, rather than the API price of a fixed amount of text.

On another OpenAI evaluation, Agents’ Last Exam V1, Sol at max effort scored 56.4%. OpenAI says this was above Claude Opus 5’s highest score in that evaluation at 60% lower cost per task. The test covers long-running professional tasks across 55 sub-industries. For builders of systems that move between applications, those tasks are closer to a whole workflow than a single prompt, though the reported scores remain results for the tested tasks and settings.

OpenAI reports that Sol made about half as many factual mistakes as GPT-5.6 Sol in its internal evaluation. It built that evaluation from de-identified ChatGPT conversations in which users had flagged a factual error by an earlier model. OpenAI says those error-inducing exchanges are not representative of typical use, where factual errors are rarer. The scores were not controlled for answer length, although the company says its checks across different response lengths found almost no relationship between length and the result.

GitHub Copilot adds both GPT-6 models

GitHub made Sol and Luna available as model choices in Copilot on 22 September. Its description places Sol with interactive coding and tasks involving multiple validation steps, and Luna with smaller, faster requests. GitHub says the models use usage-based billing, so OpenAI’s API token rates should not be treated as Copilot charges.

Sol is available on Copilot Pro+, Max, Business and Enterprise plans, while Luna is also available on Pro. GitHub lists access through its model picker in environments including Visual Studio Code, the Copilot CLI, github.com and GitHub Mobile. GitHub says the rollout will be gradual. For Business and Enterprise plans, administrators can manage access through the Copilot model policy, and new models are enabled by default unless an administrator has disabled that default or the model itself.