The model picker, finally demystified.
Most people never touch the model picker, and mostly that's fine — ChatGPT routes for you. But 'mostly fine' leaves real quality on the table. Pro+ users learn the five minutes of theory that make the picker a tool instead of a mystery.
01 The family, in one table
An earlier version of this table had Sol and Luna the wrong way round. Per OpenAI’s Help Center: Sol is the heavyweight that powers every reasoning level on paid plans; Luna is the fast, low-cost model that Free and Go users get. Terra sits between them. We’ve rewritten the section to match what the picker actually shows you.
| Model | Built for | Where you actually meet it |
|---|---|---|
| Sol | Complex work: coding, research, analysis, agents | Every reasoning level on Plus, Pro, Business, Enterprise — Instant through Extra High. Sol Pro is its top-end option. |
| Terra | Balanced capability, speed, and cost | ChatGPT Work, Codex, and the API — not selectable in a standard chat |
| Luna | Fastest and cheapest | The default for Free and Go (it also powers their “Think” button); Work, Codex, and the API |
| GPT-6 Pro (Astra) | Long, consequential, computer-driving jobs | The Pro section of the picker on Pro $100/$200, Business, and Enterprise — metered per message |
Here is the thing most explainers miss: on a paid plan, the picker in a normal chat isn’t choosing between Sol, Terra, and Luna at all. Those three are the GPT-5.6 family (they went GA alongside ChatGPT Work), but in Chat you get Sol every time; what you’re choosing is how hard Sol thinks — Instant, Medium, High, or Extra High (Extra High is Pro, Business, and Enterprise only). Terra and Luna only become a choice in Work, Codex, and the API. The separate Pro section holds the two big-ticket models, GPT-5.6 Sol Pro and, as of September 2026, GPT-6 Pro.
GPT-6 Astra landed in the picker as “GPT-6 Pro.” It is the first new generation since GPT-5, it is built for computer use and template-faithful documents, and it is rationed: 200 messages a week on Pro $200, 50 a week on Pro $100 and Business Premium, 15 a month on Business Standard, off by default for Enterprise until an admin enables it, and (per the Help Center) not in Chat for Plus. It also pauses itself when a monitor flags an action — expect a “review this” prompt on long jobs. Who gets it, the limits, and what it does differently are in GPT-6 Astra: who actually gets it. Picker rule: it belongs on the relied on rung of section 04, not the everyday one.
OpenAI o3 is gone from ChatGPT as of today, August 26, 2026. It was the last of the older reasoning models still reachable, and only by paid users digging into model settings. The useful part is the pattern behind the date: OpenAI announced this retirement on May 28 and ran a 90-day sunset, which is roughly how it handles every model it removes — successor ships, predecessor stays about 90 days, then goes. Two implications for you. First, a pinned old model is always a temporary arrangement, so learn the current family rather than a favourite. Second, this is a ChatGPT change; the API is untouched, which is why a developer you work with may still be using a model you can no longer select.
02 Thinking modes: paying with seconds
Separately from which model, you can ask for more thinking — extended reasoning before the answer. The mental model: you're buying answer quality with waiting time. A thinking response might take noticeably longer and be dramatically more careful. When to spend: math you'll rely on, gnarly logic, "find the flaw in this plan," legal-ish reading. When not to: brainstorms, drafts, anything where volume beats precision — thinking mode on a brainstorm just gets you fewer, stiffer ideas, slower.
03 When auto-routing isn't enough
ChatGPT increasingly routes hard prompts to deeper processing on its own. The auto-router's weakness: it judges the question's difficulty, not the answer's stakes. "Should I take this job offer?" reads like simple chat. It is not. Your rule as a power user: when the cost of a shallow answer is high, escalate manually — push the slider to High or Extra High, or simply add "take your time and reason carefully through the tradeoffs before answering."
Best of both: draft at Instant, then hand the result to High (or GPT-6 Pro, if you have it) with "critique this hard — what's wrong, missing, or wishful?" Fast generation, deep review. It's how good teams work, staffed by one person and a picker.
04 A working default
Take a real decision you're facing. Ask it three ways: Instant, High, then the two-pass critique. Compare what each caught. You'll never see the picker as decoration again.
Open ChatGPT →This week's challenge
For one week, before every substantial ask, spend three seconds on one question: kept, or relied on? Route accordingly. Friday, look at your week's outputs and ask whether the escalated ones earned their extra seconds. (They did.)