GPT-6 Sol is OpenAI’s new option for complex coding and tasks that use several tools or steps. GPT-6 Luna focuses on lower-cost, high-volume work, such as extracting facts, sorting text, and producing structured summaries. OpenAI keeps GPT-6 Astra as its most capable model for demanding work.

The examples below offer starting points. We haven’t run a live head-to-head test for this guide.

OpenAI documents the current details on its Sol, Luna, and Astra model pages. The suggested uses follow its model-selection guidance.

A token is a small unit of text. The context window limits how much material the model can work with in a request.

The prices above cover standard API text processing for requests with up to 272,000 input tokens. Longer prompts have higher rates. ChatGPT plans and credits have separate billing terms, as explained in OpenAI’s API model pricing and product availability guide.

Try Sol for a campaign brief, a report, a code change, or a document comparison. OpenAI’s model-selection guidance positions it for writing and coding that need analysis and care.

Give it the source material and define a finished result. Save the first answer for comparison.

Start with Luna for sorting feedback, extracting fields, or turning notes into a fixed summary format. OpenAI’s model-selection guidance recommends it for focused tasks and frequent automation.

Define the allowed labels or fields and explain what to do when information is missing.

Try Astra for a project plan with conflicting deadlines, limited staff, and several requirements. OpenAI’s model-selection guidance recommends it for demanding analysis and complex work.

Check which conflicts it resolves and which still need your attention.

OpenAI’s model availability guide currently lists GPT-6 Sol and Luna in ChatGPT Work and Codex. The standard Chat view does not offer them. Access depends on your account, workspace, client, and rollout.

In Work or Codex, open the model or Power control below the composer. OpenAI’s model controls guide explains how to use Advanced to choose a specific model and effort level when those controls appear. Record the exact model name before you run a comparison.

API users can select the IDs in the table above. The Sol, Luna, and Astra model pages also offer a Try in Playground option. Check billing and access before making calls.

Effort gives the model more or less room to work through a problem. Higher settings can take longer and use more tokens. OpenAI’s model controls guide recommends starting with an available default and adjusting it for the task.

In the API, Sol and Luna default to medium effort. Both support settings from none through max. Astra supports low through max. Product presets can differ from API defaults; compare the documented settings for Sol, Luna, and Astra.

For the exercise below, use medium across the models when available and record the setting. Keep the prompt and tools the same. This gives you a repeatable setup; each model may still use a different amount of time and computation.

This exercise tests arithmetic, attention to detail, and whether the model stays within the evidence. The business and figures are fictional. You can check every calculation yourself.

Paste invented or approved notes after this prompt:

Extract the agreed tasks from these meeting notes.

Use a table with Task, Owner, Deadline, and Supporting text. Copy the
supporting text exactly from the notes. Use "Unassigned" for a missing
owner and "No deadline stated" for a missing date.

Put ideas and suggestions that received no clear agreement in a separate
section called "Needs a decision." Flag conflicts between statements.

Check that each task has support in the notes. Pay close attention to any owner or date the model adds.

Use a test project or separate branch:

Investigate this bug using the supplied steps to reproduce it.

Make the smallest change that fixes the problem. Add a test that fails
before the fix and passes afterward. Preserve behavior outside the bug.
Ask before adding a dependency.

Run the relevant tests when tools permit. Report the files changed,
tests run, results, and remaining risks. Clearly mark checks you could
not run. Wait for approval before merging or deploying.

Check the patch and test results yourself. Record unrelated edits, missed cases, and the time you spend reviewing the change.

Using 20,000 ordinary uncached input tokens and 5,000 total billed output tokens, standard API text charges would be:

These calculations use the published Luna, Sol, and Astra rates. They hold token counts equal and exclude cache operations, tool charges, and processing-tier or regional adjustments. The sample report above will have its own token count.

Billed output includes reasoning tokens, which may be hidden from the visible answer. Check the API usage record when comparing costs. OpenAI’s reasoning guide explains how effort, extra attempts, and longer answers can change what you spend.

Also watch input length: prompts above 272,000 input tokens double input and cache rates and increase output rates by 50% for the full request. The threshold is documented on the Sol, Luna, and Astra model pages.

For routine work, compare total cost across a batch, including the time you spend fixing mistakes. A small first test can help you choose which model deserves a larger trial.

Pick one task you do each week. Save a few examples, define a correct result, and compare the models with the prompts above.

Keep the prompt, settings, and first answers together. That gives you a useful record when a new model arrives or your work changes.