Model map for Claude Code

1. The quick rule

If the task repeats, splits into pieces, or needs many attempts, start with Sonnet. If the task decides architecture, money, security, or a decision that is hard to reverse, open Opus.

2. When Sonnet is enough

• Initial code scanning and understanding structure.
• A research subagent that returns a focused summary.
• A draft solution once the direction is clear.
• Small fixes after a review that already found the problem.

3. When to open Opus

• Comparing architecture alternatives.
• A stubborn bug after several attempts in the same direction.
• Security, money, permissions, or deletion.
• The final review before merge or release.

4. Short decision table

Situation: research subagent
Model: Sonnet

Situation: initial code scan
Model: Sonnet

Situation: writing a draft solution
Model: Sonnet

Situation: comparing architecture alternatives
Model: Opus

Situation: stubborn bug after several attempts
Model: Opus

Situation: final review before merge
Model: Opus

5. Common mistakes

• Running everything on Opus and then being surprised when the quota burns out.
• Letting the same model both run the work and approve critical work.
• Switching models before you define what counts as done.
• Forgetting that review should be a different role, not a continuation of the same attempt.

6. A short prompt to start with

Task: [what needs to be done]

Split the work:
1. What Sonnet can do fast.
2. Where you need Opus to decide or review.
3. What counts as done.
4. What evidence proves the work succeeded.

Do not start executing before the model split is clear.