OpenAI's model lineup just got more interesting. GPT-6 Sol and GPT-6 Luna now have official API model pages, adding two options alongside GPT-6 Astra. The split looks familiar: Sol is positioned for complex coding and agent workflows, while Luna is aimed at focused, high-volume work.
The model names are official. The rollout story is less tidy. People are already spotting the models in Codex, but early reports describe uneven access and errors. I would treat that as a rollout in progress, not proof that every ChatGPT or Codex account can use them today.
What the official model cards say
OpenAI describes GPT-6 Sol as a model for complex coding and agentic workflows. Its published API specifications list a 1.05-million-token context window, up to 128,000 output tokens, and reasoning settings from none through max. The listed API price is $2 per million input tokens and $10 per million output tokens.
GPT-6 Luna is described as OpenAI's most efficient option for focused, high-volume tasks. Its card lists the same context and output limits, and a much lower API price: $0.10 per million input tokens and $0.50 per million output tokens. Both model cards list support for tools such as web search, code interpreter, computer use, and MCP through the Responses API. These are published API specs; they do not guarantee identical availability or behavior in each ChatGPT and Codex plan.
| Model | Official positioning | API price per 1M tokens (input / output) | Context / max output |
|---|---|---|---|
| --- | --- | ---: | ---: |
| GPT-6 Sol | Complex coding and agentic workflows | $2 / $10 | 1.05M / 128K |
| GPT-6 Luna | Focused, high-volume tasks | $0.10 / $0.50 | 1.05M / 128K |
Those numbers make Luna's intended role clear: lower-cost throughput. They do not tell us whether Luna is a good fit for a particular coding task, or how often a cheaper model will need retries. Cost per successful result matters more than cost per token.
Where are the benchmark scores?
The API model cards I checked publish capability descriptions and specifications, but no comparative benchmark table. I have not found an official GPT-6 Sol-versus-Luna scorecard to support claims that either model beats Astra, Opus, or another model on coding or reasoning. So this article does not turn pricing or context size into a performance benchmark.
The practical test is still ahead: give both models the same real task, repository, tools, and acceptance checks. Compare whether they understand the request, make a small correct change, run the right tests, and report limitations. I plan to test both models against the kind of coding and agent work I actually do, then share what holds up.
What people are saying so far
The first reaction online is a mix of excitement and confusion. Codex users are posting screenshots of Sol and Luna appearing in model selectors, while others report that a selection is unavailable or returns an error. Those are useful signs of a staged rollout, but they are anecdotes rather than a reliable availability map or an independent benchmark.
The most interesting question is not whether OpenAI added two more names. It is whether Sol can cover demanding coding sessions at a lower cost than Astra, and whether Luna is capable enough to handle everyday changes without turning review into cleanup. Model labels only become useful after repeated work.
For now, the facts are narrower: OpenAI has published the two API model cards, their intended roles, and pricing. Availability in individual ChatGPT and Codex accounts may vary, and independent task-level comparisons are still needed.
Sources
FAQ
What are GPT-6 Sol and GPT-6 Luna?
OpenAI's model cards position Sol for complex coding and agentic workflows, and Luna for focused, high-volume tasks.
Are GPT-6 Sol and Luna available to everyone in ChatGPT and Codex?
The API model pages are live, but early user reports show inconsistent access in Codex. Availability can depend on rollout stage, product, and account.
Which model has the better benchmark score?
The model cards do not include comparative benchmark scores. A fair answer needs reproducible tests on the same tasks and setup.
Will I test these models?
Yes. I plan to run both through practical coding and agent workflows and report the results after testing, not from first impressions.
