GPT-5.6 Model Selection Guide: Sol, Terra, Luna, and Reasoning Effort

Recommendation

For unfamiliar work, begin with Sol at medium effort. Move to Terra or Luna only after you have a result that can serve as a baseline.

Decision table

Choose by the shape of the work

A useful path is Sol → tighten the instructions → test Terra or Luna. Do not begin with the cheapest model and spend the saving on retries.

Standard ChatGPT uses Sol for its reasoning tier. The full three-model choice is available only in supported Codex, Work, or API experiences.

Important differences

Raise effort only for a reasoning problem

Higher effort can raise latency and usage without a matching quality gain. Change one variable at a time.

For example, classifying 1,000 rows under a clear schema is a Luna low/medium candidate. A build failure with several plausible causes belongs on Sol medium first.

Max and Ultra are not the same

Use Max for one tightly coupled problem. Use Ultra when sources, logs, test suites, or components can be examined independently and combined later.

If either option is missing, check product support and settings. Availability depends on the plan and surface.

Choose by your situation

Build a baseline before optimizing

Save a small evaluation set and record quality, latency, retries, token use, and correction time.

1. Run Sol at medium effort.
2. If cost or speed matters, change only the model.
3. If reasoning is shallow, keep the model and raise only the effort.
4. Keep the cheaper setting only when it clears the same acceptance bar.

That turns the price change into a useful option without sending ambiguous or high-consequence work to a smaller model too early.

Building the evaluation set that makes the choice for you

Twenty real tasks beat two hundred invented ones

The reason model comparisons stall is almost never the models. It is that nobody wrote down what a good answer looks like before running the test, so every result becomes a matter of impression.

Pull the tasks from work that has already happened. Real inputs carry the messiness that decides outcomes: truncated logs, contradictory requirements, the file that is four times longer than the others. Invented examples are uniformly well-formed, which is exactly the condition under which every model looks adequate.

Write the acceptance criterion beside each task before running anything. “Correct” is not a criterion; “identifies the null dereference in the third function and does not alter the public signature” is.

Measure correction time, not output quality

The number that predicts whether a setting is worth keeping is how long a person spent turning the output into something usable. It captures what token counts and subjective quality ratings both miss, and it is the figure that translates directly into whether the change was worth making.

Record it per task, in minutes, alongside the retry count. A model that is right eight times in ten but produces a plausible-looking wrong answer the other two can cost more than one that fails visibly, because a visible failure is cheap to catch and a plausible one is not.

Selection mistakes that look like model problems

Four causes to rule out before changing the model

• Changing two things at once. Moving model and effort together produces a result you cannot attribute. The comparison has to isolate one variable, which is slower and is the only way the answer means anything.
• Blaming the model for the prompt. If the instruction omits a constraint, every model will miss it, and the one that appears to satisfy it is guessing. Add the constraint and re-run before concluding anything.
• Testing in a warmed-up conversation. Context from earlier turns quietly supplies information the model would not otherwise have. Compare in fresh sessions.
• Rating fluency as correctness. Higher-effort settings tend to produce more organised prose, which reads as more competent regardless of whether the substance improved. Check the claim, not the paragraph.

The last one is the most common and the hardest to notice in yourself. It is the argument for writing acceptance criteria down in advance: a criterion fixed before the output arrives cannot be quietly relaxed by a persuasive answer.

Re-run the baseline when the platform changes

A saved evaluation set is worth keeping precisely because model lineups and prices move. When a new option appears or a price changes, the question is not whether the new setting is good but whether it clears the bar your own tasks already established. That takes an afternoon with a stored evaluation set and a week without one.

What could change the choice

Task difficulty, acceptable latency, credits, and access to Max or Ultra can change the starting point. Keep a measured baseline for comparison.

Limits and evidence

Sources were reviewed on July 31, 2026. The recommendation compares the documented options and recorded tests, not every workload, region, plan, or future release. Recheck the conditions that drive the choice before using it elsewhere.

Sources

• OpenAI: Codex models: model roles, recommended power settings, reasoning effort, Max, and Ultra. Checked July 31, 2026.
• OpenAI: Subagents: Ultra, parallel subagents, and coordination limits. Checked July 31, 2026.
• OpenAI: GPT-5.6: product availability and the July 30 pricing and credit update. Checked July 31, 2026.
• OpenAI Help Center: GPT-5.6 in ChatGPT: the reasoning tier in ChatGPT and the broader model choices in Codex and Work. Checked July 31, 2026.

Related articles

• GPT-5.6 on Amazon Bedrock: Regions, Model Choice, and Cache Economics