How to Use GPT-6 Sol and Luna in ChatGPT Work and Codex: Model Picker, Reasoning Levels and Best Use Cases

When you select a specific model in Work or Codex, the interface can expose that model’s available reasoning levels. OpenAI says the exact options vary by plan and workspace settings.

The useful mental model is:

• Lower reasoning: faster responses for routine edits, simple code changes, straightforward analysis and tasks where iteration is cheap.
• Medium/high reasoning: better fit for multi-step debugging, architecture decisions, research synthesis, complex documents and tasks where mistakes are expensive.
• Highest available reasoning: reserve for difficult problems that genuinely benefit from more deliberation; it can increase latency and usage.

Do not automatically maximize reasoning for every request. A simple rename or formatting task does not need the same effort as a cross-repository migration or multi-source research project.

OpenAI positions GPT-6 Sol as a balance of intelligence and cost.

A useful way to route work is by the cost of being wrong. For a low-stakes transformation—renaming fields, extracting a known schema or producing many first-pass summaries—throughput can matter more than maximum reasoning depth. For a migration plan, production incident, architecture decision or research conclusion that will drive a business action, the quality of the reasoning and verification process matters more. This is where Sol is easier to justify even if a lighter model could produce a plausible first answer.

Model choice also interacts with context. A short request with one clear file is different from a task that requires reading a repository, checking documentation, comparing several files and then applying edits. In the second case, the surrounding Work or Codex tool chain can be as important as the raw model. Give the model the smallest relevant set of files and permissions, define the expected output and verification step, and let it escalate effort only where the task actually needs it.

For teams, standardize a routing rule instead of relying on individual preference. One practical policy is: Luna for bounded high-volume work, Sol for multi-step professional work, and the strongest model your environment exposes only for genuinely difficult exceptions. Record the model and reasoning setting when reproducibility matters. This makes cost and quality decisions explainable later, especially when two runs produce different outcomes.

In Work and Codex, that makes Sol the sensible default for professional tasks where you want strong reasoning but do not need the maximum capability/cost profile for every step.

Typical fits include:

• Codebase changes that involve several files.
• Research and synthesis across documents/web sources.
• Technical writing and review.
• Planning and executing multi-step work with tools.
• Debugging where context and tradeoffs matter.

Sol is not automatically “better” than Luna for every job. If the task is repetitive and volume is the dominant constraint, Luna can be the more efficient choice.

Editorial context photo: Zan Lazarevic / Unsplash.