On October 2, 2026 OpenAI published an official guide to choosing between the three models in its GPT-6 family: Astra, Sol and Luna. It explains what each one is for, what it costs and what to measure before putting it to work. It is written for technical teams, but it is useful to any business that pays to use AI.
What we know
- What it is: a guide titled "A model guide for the GPT-6 family", published by OpenAI on October 2, 2026. It does not announce a new model or change prices.
- GPT-6 Astra: the most capable. The guide reserves it for the hardest reasoning work. It costs $10 per million input tokens and $50 per million output tokens.
- GPT-6.1 Sol: for complex coding, research and computer use. It costs $2 and $10. OpenAI describes it as "near-Astra intelligence" for a fifth of the price.
- GPT-6 Luna: for focused, repeated tasks with a clear goal, such as extracting invoice fields, classifying requests or writing summaries. It costs $0.10 and $0.50.
- The unit: these are API prices per million tokens, at the standard rate and with short context. The pricing page has other rates for long context and for the faster modes.
- Context window: 1.05 million tokens for all three, according to the models page.
- Caching: repeated input tokens cost up to 95% less, depending on the model.
- Reasoning effort: the guide describes four levels. Low for routine tasks, medium for work that needs judgment, high for difficult analysis, and a top level for when high is not enough.
- What to measure: success, latency and cost per successfully completed task.
What changes and what doesn't
What you pay and what is available do not change. The prices match the ones already on OpenAI's pricing page.
What changes is that OpenAI has put a rule in writing: you choose the model by the task. And the yardstick it proposes is not price per token but the cost of each task that comes out right.
The guide covers the API and Codex. It is not about ChatGPT plans: if you use ChatGPT on a subscription, nothing changes for you.
One sum to show the gap. A thousand tasks of 2,000 input tokens and 500 output tokens cost $0.45 on Luna, $9 on Sol and $45 on Astra. The sum is ours, using the published prices and no caching. The cheapest and the most expensive are a hundred times apart.
How to choose a model for your business
- Write the task down. What goes in, what must come out and when it counts as finished. The guide stresses defining what "done" means.
- Look at the kind of task. If it is repeated and has a clear goal, the guide points to Luna. If it involves coding or research, to Sol. Astra is for the hardest work.
- Test with real cases. Measure how many come out right, how long they take and what each success costs. A cheap model that fails often can end up expensive.
- Order your prompt. Instructions that never change go first and the details of each task go after. That way caching cuts the cost of what repeats.
- Tune reasoning effort. Lower it for simple work and raise it only where needed.
- Set the limits. Say what the model may do on its own and what needs a person's approval.
How we apply it at DomHostSeo
We have not benchmarked these three models against each other. The prices and recommendations are OpenAI's, read on October 2 and 3, 2026. The example sum is ours.
We use AI to draft and illustrate, within rules we wrote ourselves, and we check every fact against its primary source before publishing. We covered it in Google: Fact-Check AI Content by Hand Before Publishing.
Sources
Updates: we will add here any change to prices or models that OpenAI publishes.