Lesson 7 · 15 min
You don't call your most experienced business partner to copy a spreadsheet. Models work the same way: the light model is fast and uses less of your plan, and it handles routine work; the top model thinks more slowly and is for the big plan, the hard problem, and anything that could cost a lot. The routing rule spells this out once.
Start the lesson
Goal: write the rule that decides which model handles each task and classify your tasks as light or top.
Click each part.
When in doubt, start with the light model and have it say so if it gets stuck. Moving up is easy; using the top model for everything is what gets expensive.
Names change over time. Check your model list.
In ChatGPT, Claude, and Gemini, the model selector is at the top of the conversation or near the text box.
Each option’s description tells you which one is fast for everyday use and which one reasons more deeply or is the most capable: the first is your light model, the second your top model. In Claude, Sonnet is for everyday use, and Opus is the most capable.
To move up, switch models in the selector. Many tools let you switch in the same conversation; a new conversation in the same Project starts fresh, with only the sheet.
In Codex, /model shows the list for your plan. In the list dated Nov 4, 2026: gpt-6-luna (fast and economical, for easier tasks), gpt-6.1-terra, gpt-6.1-sol (the general-purpose one), and gpt-6-astra (for more demanding work). In practice: luna is the light model, astra is the top model; terra and sol are in the middle.
In the command, the model goes after -m.
In Claude Code, /model switches the session model, and --model selects one when you launch it: sonnet for routine work, opus for difficult tasks.
The aliases always point to the latest version in each family.
With a subscription, the top model usually uses more of your usage limit and takes longer to respond.
Routine work on the light model keeps your limit free for when you really need the top model.
In chat, add it to the Project instructions; in the terminal, add it to the end of AGENTS.md or CLAUDE.md.
Model rule: Use the LIGHT model for routine tasks: - reading dashboards, spreadsheets, and messages; copying data from one place to another - weekly reports, summaries, and tables - email drafts and short replies - small edits to a text, spreadsheet, or page Switch to the TOP model when: - it's a big plan or a new system - the light model has made the same mistake twice already - you need to DECIDE something involving money or a contract (payment, negotiation, accepting terms), or make a proposal for a major client (reading and adding up invoices or bills is routine: use the light model) When a task calls for the top model, tell me in one line: "I suggest the top model because ..." and wait for me to switch models. If you're unsure, start with the light model and let me know if you get stuck. Drafts and gates apply to both.
Use the model rule above. Here’s my task list (lesson 1). Return one line per task, in this format: N | task | light or top | reason in a few words
# routine work on the light model codex exec --skip-git-repo-check -m gpt-6-luna -s read-only -o report.md "report: follow AGENTS.md" claude -p --model sonnet "report: follow CLAUDE.md" > report.md # difficult work on the top model codex exec --skip-git-repo-check -m gpt-6-astra -s read-only -o plan.md "build a plan for the new estimating process" claude -p --model opus "build a plan for the new estimating process" > plan.md
1 | Weekly website report | light | read and summarize 2 | New contacts and drafts | light | routine, with a gate 3 | Invoices and bills | light | read and add up; top if the PDF can't be read 4 | New client checklist | light | fixed steps 5 | Design the new quoting process | top | new system
Paste the list the agent returned (or your own). The page checks that each line has just one level, light or top. You decide whether the level is right by asking the three questions.
One rule, five tasks, one test.
Paste the routing rule into the instructions for your agents’ Projects (or at the end of AGENTS.md / CLAUDE.md).
Done when: when you ask "which model would you use for the weekly report?", the agent answers light and gives the reason.
Use the "classify my tasks" request with the list from lesson 1. Check it with the checker.
Done when: the checker responds OK, and every task that DECIDES money or a contract is marked top or "do not delegate"; reading and adding up invoices can stay with the light model.
Open the model selector in the tool you use (or /model in the terminal), read each option’s description, and write down the name of the light model (the fast one for everyday use) and the top model (the one that reasons more).
Done when: you have both names written down and ran the weekly report on the light model.
The explanation appears after your answer.
Question 1 of 4
Which of these tasks should go to the top model?
Question 2 of 4
The light model made the same calculation error twice in the report. According to the rule:
Question 3 of 4
Why not use the top model for everything, since it's better?
Question 4 of 4
In chat, how do you send a difficult task to the top model?
Choose an option. The explanation appears after you try.