Skip to main content
OpenAI expanded the GPT-6 line with two models for everyday work: GPT-6 Sol and GPT-6 Luna. Both were trained with methods similar to GPT-6 Astra and bring that step to coding, automation, computer use, and professional work. Astra remains OpenAI’s model for the most demanding work. Sol covers difficult professional tasks. Luna is the model for high volume and focused tasks. OpenAI built-in tools, such as web search, code interpreter, computer use, and MCP, go through the Responses API.

​
What changes from GPT-5.6

OpenAI also brought Astra’s communication style to Sol and Luna: less jargon, fewer low-value details, and slightly shorter answers, especially in technical conversations.

​
What the benchmarks show

The numbers below come from OpenAI’s announcement. The level in parentheses is the one used for that measurement. On OpenAI’s internal factuality evaluation, Sol makes about half as many mistakes as its predecessor. At the higher levels, Luna approaches GPT-5.6 Sol.
These results were published by OpenAI and depend on the harness, tools, and the level used in the measurement. The AutomationBench comparison with Fable 5.1 does not include the Opus 5 fallback, which happened on about 40% of tasks. Validate the model in your own workflow before replacing a production configuration.

​
Where Astra stays ahead

Sol and Luna do not replace Astra on the hardest work. OpenAI keeps Astra as the best model in the family, including for computer use. Use Sol or Luna when Astra is more than the task needs.

​
Where each one creates more value

​
Sol, when the task is hard and will still iterate

Multi-file coding, automation across apps, and professional analysis the team will review and revise. Expected result: a reviewable change or analysis, with an explicit check of what was verified.

​
Luna, when volume is the problem

Classification, extraction, triage, and agents that repeat the same kind of task all day. At the higher levels, OpenAI positions Luna near larger models on coding and computer use. Expected result: the same operation, at scale.

​
Cache, when the agent reuses the same context

Long sessions and agents that resend instructions, files, and history reuse more of the prefix. Tools can change mid-conversation without losing what was already processed. Expected result: faster responses in sessions that repeat the same context.

​
How to apply it in Tess

1

Pick the model for the work

Luna for volume and focused tasks. Sol for coding and difficult professional workflows. Astra when Sol’s quality still does not close the deliverable.
2

Provide context and limit tools

Attach only the material the task needs and enable the tools it actually uses. Ask for confirmation before publishing, sending, or changing production.
3

Ask for a check before finishing

Tell the agent to verify tests, sources, and numbers, and to say what was not verified.

​
Two playbooks to start

​
1. A code change with Sol

Input: repository, expected behavior, and test commands.
Done when: the change is in the repository and the relevant tests pass.

​
2. High-volume triage with Luna

Input: a queue of items, the classification rule, and the output format.
Done when: every item has a traceable label and ambiguous cases are separated for human review.

​
Limitations

  • A 1.05M context window does not replace selection. Send only the relevant material.
  • OpenAI built-in tools go through the Responses API.
  • Sol and Luna are not in OpenAI Chat yet. In the API, the IDs are gpt-6-sol and gpt-6-luna.
  • Review code, numbers, and citations before publishing. Keep human approval for irreversible actions.
See also: Models and Costs · GPT-6 Astra · Introducing GPT-6 Sol and Luna · GPT-6 Sol · GPT-6 Luna.