Guide · reference

Which model, and when

Claude Code and Codex each come with several models. A model is the brain behind the agent. The big ones think better and cost more of your allowance. The small ones are quick and cheap. You pick with one command, and picking well is the difference between running out at lunchtime and building all day.

You need

Claude Code (chapter 2) or Codex (chapter 3), logged in.

The one rule

Think with the big model. Type with the small one.

Plan with the strongest model your plan gives you. Planning is a few hundred words of thinking, so it costs little, and a good plan saves hours.
Switch down to do the work. Carrying out a clear plan, step by step, is what the everyday model is for. It is quicker and it uses far less of your allowance.
Switch up when you are stuck. A bug the small model has failed on twice, a review before you call something done, a design choice you cannot undo. Those are worth the big one.

Both agents switch with /model. Type it, pick from the list, carry on. The conversation stays. Many people never change the default, and that is fine too. The default is chosen to be the sensible middle.

The models

Biggest first. Checked on 2026-11-02; names change a few times a year.

Claude Code

Fable 5.1. The strongest. For planning a whole project, the hardest bugs, and a last look before you ship. On Max and Ultra you can spend up to half your weekly limit on it. On Pro it is not in your plan and costs extra, so Pro readers plan with Opus instead.
Opus 5.5. The big model for most people. Plans, reviews, anything you have failed at twice. On Pro and up.
Sonnet 5.5. The everyday builder. Carrying out a plan, adding a feature, fixing a clear error. Most of your day should be here.
Haiku 5.5. Small and quick. Renaming, tidying, writing a commit message, helper jobs. Rarely worth picking by hand, but the agent uses it behind the scenes.

Claude Code also has a Fast mode, which runs Opus with faster output. Same brain, quicker answers, and it uses more of your allowance. Try it when you are waiting a lot.

Fable does not show up in old versions. If it is missing from /model, update: Fable 5.1 needs Claude Code 2.1.255 or newer. Free accounts cannot use Claude Code at all.

Codex

GPT-6 Astra. The strongest. OpenAI calls it the model for "complex work across code, apps, and research". On Plus it is expensive: roughly 5 to 45 messages in a 5-hour window, so save it for plans, reviews and the hard bugs.
GPT-6.1 Sol. "Near-Astra performance for complex work at a lower cost." The everyday builder on Plus and Pro, about three times as many messages as Astra.
GPT-6.1 Terra. For writing, understanding images and screenshots, and driving a computer. Needs Codex 0.159 or newer.
GPT-6 Luna. Small and quick, for repeatable jobs like extracting, sorting and summarising. The only model on the Free and Go plans, in the desktop app.

Codex also lets you set reasoning effort, how hard the model thinks before it answers: Light (shown as Low in the terminal), Medium, High, Extra High, Max and Ultra. Pick it under /model. OpenAI's own advice: start with the default, raise it for work that needs more planning or checking, and "most tasks do not need Max or Ultra". Astra can start on Light. Luna wants High.

ChatGPT Pro plans have no 5-hour limit, only weekly ones. Plus has both. /status shows where you are.

What it costs you

Neither company tells you an exact price per message on a subscription. What you can count on:

The big model uses several times the allowance of the everyday one, per message. On Codex Plus, OpenAI's own estimate is about 5 to 45 Astra messages per 5 hours against 15 to 160 for GPT-6.1 Sol and 350 to 3,000 for Luna.
Long conversations cost more than short ones on every model, because the agent rereads everything each turn. /clear between tasks matters more than which model you pick.
When you hit a limit, nothing is lost. Switch to a smaller model and keep going, or wait for the reset.

A day with the rule

Morning. Switch to Opus, Fable or Astra. Paste the project brief. Ask for a plan (chapter 4, rule 3). Read it. Argue with it.
Switch to Sonnet or GPT-6.1 Sol. Work through the plan one step at a time, looking after each.
Something will not fix after two tries. Switch up, paste the exact error and what you tried, let it look. Switch back down.
Before you push, switch up once more: "Review the changes since this morning. What would you fix before shipping?"

Helpers, when you are ready

Claude Code can run small helper agents, each on a model you choose, so the big one plans and the cheap ones do the legwork. It is a real saving on big projects and a distraction on small ones. Build three projects first. Then read about subagents in Going further.

If it didn't

What you just did

You met the models behind each agent and what your plan gets. You have one rule that keeps your allowance for the moments it matters: think with the big model, type with the small one.