GPT-5.6: Which Model and Effort Should You Actually Use?

GPT-5.6 arrived as three models and several effort settings.

That creates the wrong instinct: pick the biggest model, turn the effort up, and assume the result will be better.

After using the family, I have landed somewhere simpler.

Terra high should do most work. Sol high handles hard work. Luna belongs inside a larger workflow. Max is mostly wasted money, and Ultra remains an open question.

The real story of GPT-5.6 is that less is often more.

The everyday decision starts with Terra

OpenAI positions Terra as the balanced everyday tier, and that description is accurate.

For well-described features, reproducible bugs, bounded refactors, tests, or small reviews, I start with Terra high.

Most repository work is not an intelligence contest. The requirements are known and the result can be tested. What matters is whether the agent follows the contract.

This is where GPT-5.6 feels better than GPT-5.5. Routine work needs less steering, so the efficiency improvement is visible in the session.

Terra still needs a good question. A newer model does not turn an ambiguous prompt into a specification.

Sol earns its place when judgment matters

Architecture, cross-cutting changes, difficult debugging, security review, and conflicting evidence all require judgment about what the problem actually is. That is when I move to Sol high.

Sol is also excellent at computer use: inspecting interfaces, operating tools, checking results, and continuing through several steps. Theo Browne highlighted the same strength, while Pietro Schirano praised Sol's speed and creativity.

If I know a task is difficult, I would rather begin with Sol high than force Terra through heavier effort settings. The model upgrade is justified because judgment, not execution, is now the bottleneck.

Luna works best behind the scenes

Luna should not be the main agent for a large repository task.

It makes sense as a tool for narrow work: sorting data, extracting fields, formatting a schema, classifying a list, summarising logs, or performing an exact rename.

Luna high is enough for most of that work. Xhigh can add reliability, but it sits close enough to Terra high that Terra usually becomes the better choice unless the task remains extremely bounded.

The rule is simple: Luna needs narrow inputs, a defined output, a stopping condition, and an independent check. If the task requires architecture, broad context, or interpretation, give it back to Terra or Sol.

A failed run does not mean “use max”

Max can at least double the cost without producing a noticeable improvement. More reasoning often creates a longer route to the same answer.

When a strong run fails, I now ask:

• did I ask the right question?
• did I describe the outcome rather than my guessed solution?
• is important context missing?
• are the acceptance criteria clear?
• is the agent working at the wrong layer?

Reframing the task is usually more valuable than moving from xhigh to max.

Ultra is different. OpenAI describes it as a subagent workflow. Parallel work could help, but coordination could consume the gain. Until it beats one focused Sol run, I consider Ultra an experiment.

Frontend shows why model choice is not universal

Claude Fable 5 still has better unprompted frontend taste for me. Give both models “make this dashboard look good” and Fable is more likely to supply a coherent visual opinion.

Sol becomes excellent when the brief specifies hierarchy, typography, motion, responsive behaviour, references, and what to avoid.

The external evidence has the same split. Framer's CanvasBench found Sol strong on responsive navigation and interactions, while Fable led consistency-oriented areas. Claire Vo's comparison preferred Sol overall but kept Claude for agentic voice.

There is no universal winner because taste, tool use, judgment, and mechanical execution are different jobs.

Where I landed

I start routine work on Terra high and difficult work on Sol high. I let Luna high handle only simple, verifiable subtasks. I use Luna xhigh selectively, Sol xhigh when a well-framed problem genuinely needs more reasoning, and max almost never.

If a run fails, I fix the question before increasing the effort.

That is the useful change in GPT-5.6. Terra is finally a credible default, Sol is a deliberate escalation, and Luna is a component rather than a compromise.

The best setting is not the largest one. It is the smallest one that can do the job properly.