OpenAI released GPT-5.6 on 9 July as a family rather than a single model. Sol was the flagship, Terra balanced capability and cost, and Luna was the fastest and least expensive option. The family became available across ChatGPT, Codex, and the API.
Three models, three practical roles
The naming encouraged a useful question: how much capability does this task actually need? A routine classification or rewrite may not benefit from the largest model, while a difficult coding or research workflow may cost less overall if a stronger model avoids repeated failure.
Fast, inexpensive work with simple checks.
Everyday work needing more capability.
Difficult tasks where higher quality changes the outcome.
What you can try
Create a small routing table based on evidence. Start a task with the lowest tier likely to meet its quality bar, then escalate only when the result fails a defined check. Keep high effort and multi-agent modes for work that can justify the additional cost and supervision.
Separate routine, complex, and high-impact work.
Define what a passing result looks like.
Use the lowest suitable tier.
Move up only when the check fails.
Key takeaways
- The routing rule is documented.
- High-impact tasks keep human approval.
- Actual cost includes tools and retries.
- Model changes are tested before rollout.
GPT-5.6 widened the choice available to users. The mature response was not to send everything to Sol, but to match model, effort, permissions, and review to the value and risk of the job.
