OpenAI introduced GPT-6 Sol and GPT-6 Luna on September 22, expanding its model family below GPT-6 Astra. The release focuses on lowering the cost of repeated professional, coding and agent tasks, according to the company’s announcement.
The published standard API prices are $2 per million input tokens and $10 per million output tokens for Sol. Luna is listed at $0.10 for input and $0.50 for output. Those figures describe token pricing, rather than the total cost of a complete workflow.
Availability and the role of each model
OpenAI says the models are available through its API as gpt-6-sol and gpt-6-luna. At launch, it also announced access in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. The announcement distinguishes those surfaces from Chat, where the models were not yet available at launch.
The company positions Sol for general professional and coding work, with Luna providing a lower-cost option. Astra remains the top model in the family under OpenAI’s own positioning.
The performance comparisons in the announcement are provider-reported evaluations. They do not establish that either model will complete a particular customer’s task at the same cost or accuracy.
Caching changes the economics of repeated context
OpenAI also describes improvements to prompt caching, including more control over cached prefixes and the ability to adjust reasoning effort or tool availability without losing earlier cache reuse.
That matters for agents that repeatedly send substantial instructions or conversation history. A lower rate for cached input can reduce part of the bill, but tools, output and repeated attempts still contribute to the cost of a task.
A workflow that uses a cheaper model but needs more retries may not produce the expected saving. The useful comparison is the cost of a completed, verified result.
What to evaluate before switching
For teams already using a previous model, the release creates a reason to rerun a representative task set. It is worth checking difficult examples, structured output and tool use as well as straightforward answers.
The published prices are a clear change. Claims about better reliability and coding performance need to be assessed against the work the model will actually receive. See our guide to agents and chatbots for the additional responsibilities that appear when a model can take actions.
