OpenAI Publishes GPT-6 Model Selection Guide: Luna for Repetitive Work, Astra for Hard Problems
The guide points users to Astra for hard problems, Sol for complex coding and research, and Luna for repetitive tasks. Options that speed up responses come with an extra charge.
On October 2 (local time), OpenAI published a guide explaining how to use its GPT-6 family of models for work and in products. The document lays out how to pick the right model for the task at hand and how to write instructions for it. It also covers how to manage jobs that run for anywhere from a few hours to several days, and how to wire the models into live services.
Which model fits my work?
OpenAI sorts the models by the kind of work you hand them. GPT-6 Astra is meant for the hardest problems. GPT-6.1 Sol is described as the right fit for complex coding, research, and computer use. GPT-6 Luna is recommended for pulling out information, classifying it, and producing summaries in a fixed format. In short, Luna is the pick for repetitive jobs with a clearly defined goal.
Developers plugging the models into a service can dial up or down how much effort the AI puts into solving a problem. OpenAI suggests trying the higher settings when a model supports them. The ability to operate a computer directly is available on Astra, Sol, and Luna alike. You also need to check whether the specific feature you want is supported on the model you have in mind.
What does speed cost?
Options that make responses faster carry an extra charge. For developers, Fast mode has a higher cost per token than Standard. Tokens are the unit used to calculate billing. Ultrafast, available in Codex and in the developer-facing services, is another speed option that costs extra. OpenAI says Ultrafast can be used with GPT-6 Astra.
The guide also covers a way to cut costs when you reuse the same input over and over. Reusing stored input can be up to 95% cheaper than regular input, depending on the model, according to OpenAI. Per-model pricing has to be checked against the separately published price list.
Can I change course on a long-running job?
OpenAI says the GPT-6 family can carry out work that stretches across hours or even days. A developer feature that splits a job across multiple AI instances is available in beta on GPT-6.1 Sol. There are limits on changing instructions midstream: new instructions will not cancel tools already running or undo completed work.
As a case study, OpenAI pointed to the video editing service Invideo. Invideo reported that its success rate on color correction work was roughly three times as high. Some editors, the company said, produced about 50 effects in a single day.
The thrust of this guide is choosing the model, the speed option, and the operating setup that match the job you are handing over. Users assigning repetitive tasks will want to look closely at input costs, while those handing off long-running work should check how much room they have to change instructions along the way. Anyone planning to use these models for work in Korea should also confirm local availability, Korean-language support, eligibility requirements for each feature, and actual pricing.