We’ve mostly said use the best model — and for high-value one-offs, that’s probably a fine default. But your coding agent isn’t one task. It’s many: planning, grinding out boilerplate, hunting a nasty bug, running commands. You might not need the biggest frontier model and it might even be slower.
The best cheat sheet is already in your terminal. Run /subagents in Copilot CLI and read the default model each subagent ships with (they are likely the default ones, if you haven’t added your own):
- explore, task → Haiku — fast + cheap for high-volume grinding (search, running commands)
- research → Sonnet — a step up where synthesis matters
- rubber-duck → a different model family — a genuinely diverse second opinion
- code-review, security-review → inherit your strong driver model
That’s the Copilot team’s own task→model map — and notice it’s not “always the biggest.” The principle generalizes: a frontier model (Opus, GPT-5.6 Sol) for deep reasoning and debugging; a fast, cheap one (Haiku, GPT-5.6 Luna) for mechanical work. As one practitioner puts it: “Use Opus to think and Sonnet to build.”
Switch your own model in Copilot with /model. Claude Code has the same idea.
💡 Try this: Run /subagents (in Copilot CLI) and read the defaults — cheap+fast for grinding, a different family for critique, your strong model for review.
đź”— GitHub model comparison
🔗 Claude — choosing a model
How do you pick a model?
🟢 I match it to the task
🟡 I use the default / auto
đź”´ Always the best I can get
⚪ I don’t think too much about / I didn’t know I could change
What’s your go-to model for your main agent? Why?