GPT-5.6 Sol vs Terra vs Luna: My Early Model Selection Framework
After two days of heavy coding and switching between Sol, Terra, and Luna, here's my current model selection framework. The goal: use the strongest model when it actually matters, without burning through Codex limits on normal tasks.
This is my actual setup:
Commits, renaming, spacing, tiny UI changes: Luna Medium/High or GPT-5.4
Normal bug fix or clearly scoped feature: Luna XHigh
Unclear task that requires exploring several parts of the repo: Terra Medium
Complex bug, architecture, auth, payments, migrations: Sol Medium
Terra/Sol Medium failed: Sol High/Max
Sol Ultra: Basically never
My Workflow: Luna XHigh → Terra Medium → Sol Medium
The actual sequence I run is simple: start at Luna XHigh. If it doesn't work, step up. Don't start at the top.
This matters because Sol burns tokens ridiculously fast. I've been watching my Codex limits while running these comparisons, and the difference in consumption between Sol Medium and Luna XHigh is not subtle. Sol gets used when the task is genuinely difficult or when a bad implementation could cause serious problems. For everything else: Luna.
What I've Learned About Each
Sol is strong. Sol is expensive. The benchmark numbers are real — Sol is the best model in the lineup for hard problems. But "strong" and "use it for everything" are not the same thing. On clearly scoped tasks, the difference between Sol and Luna XHigh is negligible. On simple tasks, it's imperceptible. The tokens Sol burns aren't worth it for tasks that Luna handles fine.
Terra uses more limits than I expected. This was the surprise. Terra sits between Luna and Sol in the capability hierarchy, so I assumed it would sit between them in cost-per-task. It doesn't always work that way. On some tasks, Terra feels like it burns noticeably more than GPT-5.5. For unclear tasks that require exploration — where Terra's value is real — it's worth it. As a default model for clearly scoped work: not any more. Luna XHigh handles clearly scoped tasks at a fraction of the cost.
Luna is the everyday workhorse. Luna Medium/High for trivial changes. Luna XHigh for everything else that has a clear scope. It's cheaper, it's faster, and for the vast majority of real coding work — the commits, the bug fixes, the scoped features — it's genuinely good enough. More than good enough.
The Framework in Practice
The model selection isn't complicated. It has to do two things: match the difficulty of the task, and preserve the harder models for when they actually earn their cost.
The easy test: "Could a bad implementation of this cause real problems?" If yes — architecture decisions, auth flows, payment logic, migrations — Sol Medium minimum, Sol High/Max if the stakes are high.
If no — clearly scoped bug fix, a feature with a clear brief, a refactor you understand — Luna XHigh. It will surprise you how often that's enough.
The unclear case — "I need to explore several parts of this repo to understand what's happening" — is where Terra Medium earns its place. It has enough reasoning capability to handle ambiguity without burning Sol-level tokens.
What This Means For Your Limits
If you're on a Codex plan with usage limits, model selection isn't just about getting the best output. It's about how many tasks you can run before your window resets.
Running Sol Medium as your default means fewer tasks run before you hit the wall. Running Luna XHigh as your default means your limits last longer while getting comparable results on clearly scoped work.
The framework: be intentional about Sol. Save it for the cases where its capability actually shows up in the output. Let Luna and Terra handle the rest.
The Honest Caveat
Two days is early. My model preferences will almost certainly evolve as I run more tasks and get a better sense of where each model actually earns its cost in real usage versus benchmarks. This is my current setup — not a final verdict.
What I am confident about: the Luna XHigh → Terra Medium → Sol Medium workflow is the right starting framework. Start low, step up when needed. Don't burn Sol tokens on tasks that Luna handles fine.
That's the whole game.