model-router

name

model-router

description

Recommends the most cost-effective Claude model (Haiku 4.5 vs Sonnet 4.6) for Claude Code tasks. Use this skill whenever the user asks "which model should I use", mentions "haiku" or "sonnet" in a routing context, asks about Claude Code cost optimization, wants to save money on AI usage, says "is this a haiku task", describes a task they're about to delegate to Claude Code and seems to want guidance, or asks "can haiku handle this". Also trigger when the user pastes a claude command and wants to know if they should downgrade/upgrade the model. Trigger proactively when someone describes a Claude Code task — suggest the optimal model even if they didn't explicitly ask.

Model Router

Help the user pick the most cost-effective Claude model for their Claude Code task. The goal: spend Sonnet budget only when a task genuinely needs judgment. Mechanical tasks belong on Haiku.

Model Reference

Model	ID	Input	Output
Haiku 4.5	claude-haiku-4-5-20251001	$0.80/MTok	$4.00/MTok
Sonnet 4.6	claude-sonnet-4-6	$3.00/MTok	$15.00/MTok

Sonnet costs ~3.75× on input and ~3.75× on output. For a typical task, Haiku saves roughly 73%.

Tip: for interactive sessions, switch inside Claude Code with /model claude-haiku-4-5-20251001.

Routing Logic

The single deciding question: Does this task require judgment, or just execution?

Execution = following clear instructions mechanically. Judgment = understanding intent, weighing tradeoffs, reasoning about correctness.

Haiku — mechanical execution

No ambiguity about what "correct" looks like:

• Run tests, check CI, verify builds pass
• Format, lint, or auto-fix code (Prettier, ESLint --fix)
• Git operations: commit, push, status, diff, log
• Install or update dependencies
• Search / grep / find across the codebase
• Rename identifiers or files across the codebase (find-and-replace level)
• Generate boilerplate from a clear template
• Convert between simple data formats (JSON ↔ CSV, etc.)
• Add a logging statement or debug print
• Read a file and return a specific piece of information
• Run a script and report what happened

Sonnet — judgment required

The task requires understanding intent, context, tradeoffs, or correctness:

• Code review: spotting bugs, evaluating design, suggesting improvements
• Debugging a mysterious failure or unexpected behavior
• Designing or architecting a new feature
• Implementing a non-trivial feature from requirements
• Refactoring for maintainability or correctness
• Performance analysis and optimization
• Security review or vulnerability assessment
• Writing documentation that requires understanding the codebase
• Any task where a wrong answer could silently introduce a bug
• Multi-file reasoning across unfamiliar code

Borderline cases

When uncertain, ask: "If Haiku got this slightly wrong, would I catch it in review?"

• Yes, low blast radius → Haiku. You'll see it before it ships.
• No, silent failure risk → Sonnet. The model upgrade costs less than the bug.

Borderline task	Haiku if…	Sonnet if…
Add error handling	simple null-check or try/catch	needs domain knowledge of which errors matter
Write unit tests	straightforward happy-path tests	edge cases need business logic understanding
Update a config file	key/value is obvious	config has subtle interactions
Translate code to another language	syntax-level translation	idioms and patterns need to carry over

Output Format

Always respond with this structure:

**Recommendation: [Haiku 4.5 / Sonnet 4.6]**

[1-2 sentences: what makes this task mechanical vs judgment-requiring]

**Cost:** Haiku ~$0.010 vs Sonnet ~$0.039 per typical 5K-token task (~74% savings with Haiku)

**Command:**
claude --model [model-id] '[task description]'

For ambiguous tasks, add one of:

• "If you want safety: use Sonnet — the quality gap is worth it here."
• "If cost matters: Haiku is likely fine — you'll catch any issues in review."

Cost Quick Reference

For a typical 5K-token task (3K input + 2K output):

• Haiku: $0.0024 + $0.0080 = ~$0.010
• Sonnet: $0.009 + $0.030 = ~$0.039
• Savings with Haiku: ~74%

For a larger 30K-token task (20K in + 10K out):

• Haiku: $0.016 + $0.040 = ~$0.056
• Sonnet: $0.060 + $0.150 = ~$0.210
• Savings with Haiku: ~73%

The ratio is stable: Haiku consistently runs at ~27% the cost of Sonnet.