## Definition
A **model selection strategy** maps tasks to the right tier in a model fleet — typically Haiku, Sonnet, Opus — instead of defaulting one model for everything.
## The Three-Role Rubric
- **Sonnet — the Builder.** Default for *implementation*. Fast enough to keep flow, cheap enough to run all day. The right answer when in doubt.
- **Opus — the Architect.** Reserve for *decisions*. Slower and more expensive, dramatically better at cross-file synthesis and "should we even do this" questions.
- **Haiku — the Scout.** Built for *exploration and filtering*. Fan-out searches, classifications, hook-driven routing.
## Selection Heuristics
| Pick… | When… |
| ------ | ------------------------------------------------------------------------ |
| Sonnet | Clear spec, 1–5 files, streaming flow matters |
| Opus | Cross-file architecture, public API, irreversible call |
| Haiku | Large corpus filtering, parallel fan-out, frequent cheap classifications |
## Budget Heuristic — 90/9/1
A working starting point for a feature-shipping week:
- ~90% Sonnet (the actual coding).
- ~9% Opus (planning, review, gnarly bugs).
- ~1% Haiku (scouts, hooks, classification).
Opus drifting above ~20% is *over-thinking, under-shipping*; below ~3% is *under-thinking, shipping bugs*.
## Configuration (Claude Code)
```jsonc
{
"model": "claude-sonnet-4-6",
"subagents": {
"architect": "claude-opus-4-7",
"scout": "claude-haiku-4-5-20251001"
}
}
```
## Related
- [[Reasoning Budget]]
- [[Specialized Agent]]
- [[Prompt Caching]]