$ recombobulate _
home / tips / switch-models-mid-conversation-with-model
0

Switch Models Mid-Conversation with /model

bagwaa @bagwaa · Mar 25, 2026 · Performance
switch-models-mid-conversation-with-model

Not every task needs the most powerful (and expensive) model. Switch on the fly with /model.

/model sonnet

This is a game-changer for cost management. Use Opus for complex architecture decisions and tricky debugging, then drop to Sonnet for straightforward file edits and boilerplate generation.

A practical workflow:

# Start with Opus for the hard thinking
/model opus
"Design the database schema for a multi-tenant SaaS app"

# Switch to Sonnet for the implementation grunt work
/model sonnet
"Now create the migration files based on that schema"

You can also set the default model when launching Claude Code:

claude --model sonnet

Here's a rough guide for when to use each:

  • Opus — complex debugging, architecture decisions, multi-file refactors, subtle logic bugs
  • Sonnet — boilerplate generation, simple edits, test writing, documentation, routine tasks
  • Haiku — quick questions, formatting, simple lookups

The context window carries over when you switch, so Claude doesn't lose any conversation history.

Match the model to the task — your wallet will thank you.

~/recombobulate $ tip --comments --count=0

Log in to leave a comment.

~/recombobulate $ tip --related --limit=3
0
Filter Test Output with a PreToolUse Hook to Cut Token Costs

A PreToolUse hook can intercept test runner commands and filter output to show only failures, cutting thousands of tokens from Claude's context.

bagwaa @bagwaa · 3 hours ago
0
Move Specialised CLAUDE.md Instructions into Skills to Shrink Context

CLAUDE.md loads into every message. Move workflow-specific instructions into skills that load on demand to reduce token costs across your session.

bagwaa @bagwaa · 3 hours ago
0
Use prompt.id to Trace All Activity from a Single User Prompt

Every event emitted while processing a single prompt shares a prompt.id UUID, letting you trace the complete chain of API calls and tool executions.

bagwaa @bagwaa · 3 hours ago