OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, offering developers lower-priced alternatives to its flagship GPT-6 Astra. Sol is designed for complex coding and agents that carry out multistep tasks; Luna targets narrower tasks run at high volume.
Sol's standard API rates are $2 per million input tokens and $10 per million output tokens. Luna costs $0.10 and $0.50, respectively. Tokens are the chunks of text a model reads or generates. Longer answers, repeated attempts and tool calls can therefore change the cost of finishing a job, even at the same token rate.
Both models have a 1.05-million-token context window, the amount of material they can work with in a request. Prompts above 272,000 input tokens carry higher rates. The API names are gpt-6-sol and gpt-6-luna.
At launch, OpenAI said paid users could access the models in Codex and ChatGPT Work, with access rolling out gradually. Free and Go users could use Luna in the desktop app.
How Sol and Luna performed in OpenAI's tests
Sol scored 68.8% and Luna 66.6% on DeepSWE 1.1, which tests extended software-engineering tasks in real codebases. OpenAI ran both at maximum reasoning effort. Those scores do not establish how the models compare at lower effort settings or on a different team's code.
In a separate internal test, Sol made roughly half as many factual mistakes as GPT-5.6 Sol, according to OpenAI. The test used conversations in which users had flagged errors from earlier models. It measures performance on those difficult examples, not the error rate someone should expect in everyday use.
OpenAI ran its evaluations in a research environment or through the API. Different prompts and tools in ChatGPT can produce different results.
