GPT-6 Sol and Luna: who gets them and what they cost

OpenAI has added GPT-6 Sol and GPT-6 Luna to its model range, following Astra earlier this month. The launch gives people using its work and coding tools two more options, while developers get lower token prices. OpenAI’s dated announcement was posted on 22 September 2026.

First, check which part of ChatGPT you are using

The distinction between Work, Codex and Chat is easy to miss in a launch headline. If you open a normal chat and cannot find the models, that does not contradict the announced rollout. Check the product named in the release before changing your plan to chase access.

OpenAI announced a gradual rollout. Availability in someone else’s screenshot is not a reliable guide to what your own account currently offers. This article describes the published access rules, rather than claiming we have checked every plan or Australian account.

Background: our coverage of GPT-6 Astra

Sol and Luna have very different API prices

Sol’s model documentation positions it for complex coding and agent workflows. Its standard rates are US$2 per million input tokens, US$0.20 for cached input and US$10 per million output tokens.

Luna is positioned for focused, high-volume tasks. Its corresponding standard rates are US$0.10 for input, US$0.01 for cached input and US$0.50 for output, each per million tokens.

Input is the content an application sends to the model; output is what the model generates. Cached input refers to reusable material that qualifies for the cached rate. Do not assume every repeated prompt will be billed that way.

These amounts are US dollars: OpenAI’s billing help says API credit purchases remain limited to USD. They are not the monthly price of a ChatGPT plan.

For an illustrative request using 100,000 uncached input tokens and 10,000 output tokens, the standard token cost is US$0.30 for Sol or US$0.015 for Luna. That is 30 US cents versus 1.5 US cents, before tools and other applicable charges. It compares an identical token count, not identical answer quality.

There are pricing conditions. Both model pages say prompts above 272,000 input tokens attract higher rates for the full request: twice the input and cache rates, and 1.5 times the output rate. Processing options can also change the bill. Use the live price documentation when budgeting an application.

The “50% cheaper” headline needs one qualification

OpenAI compares the new rates with GPT-5.6 promotional pricing. Sol’s listed input and output rates each halve. Luna’s input rate also halves, but its listed output rate moves from US$1.20 to US$0.50 per million tokens: a reduction of about 58.3%.

The individual prices are more useful than treating 50% as a universal discount. Neither the comparison nor the worked example establishes a reduction in your ChatGPT subscription bill.

A sensible way to choose between them

Start with a small sample of the actual work. For extracting dates and amounts from documents, include a cancelled booking, a revised invoice and a file with missing information. Those cases reveal more than ten tidy examples that all follow the same template.

Record incorrect answers and the time spent checking them, as well as the API charge. If the cheaper model needs repeated correction, the lowest token price may not produce the cheapest finished job. Conversely, do not pay more for a task that the cheaper option already completes reliably. These are evaluation suggestions, not results from our own benchmark.

OpenAI reports factuality improvements, but its evaluation uses conversations selected because earlier models made mistakes. The company explicitly says that sample is not representative of typical use. It cannot tell you the chance of an error in your next answer.

For subscribers, the immediate question is whether either model improves a task you already do. For developers, it is whether the quality holds up at the new price. In both cases, keep the original material beside the answer until you know what you can trust.