save-money: The Claude Cost Router That Pays For Itself
TL;DR
Auto-detect task complexity for Claude models and route to the cheapest one that gets the job done. Simple tasks → Haiku. Complex tasks → Sonnet. Save 50%+ on API bills without touching output quality.
10-Second Pitch
- What it does: Inspects each incoming prompt, classifies complexity, dispatches to Haiku (fast/cheap) or Sonnet (capable/pricey)
- Key win: Haiku is ~20x cheaper than Sonnet for simple tasks; this skill routes automatically
- Best for: High-volume AI products where every cent counts
Setup
bash pip install claude-router ` Route a prompt with a single call: `bash claude-router route "Explain quantum entanglement in one sentence" claude-router route "Write a comprehensive technical spec for a distributed key-value store"
How It Works
1. Classify — The router reads your prompt and scores it for complexity (factual, procedural, creative, analytical) 2. Route — Simple prompts (factual Q&A, greetings, lookups) go to Haiku. Complex (multi-step reasoning, code generation, analysis) go to Sonnet 3. Respond — You get the response. The routing is transparent — you can audit which model handled each request
Pros / Cons
| Pros | Cons |
|---|---|
| 50%+ cost reduction on typical workloads | Requires accurate complexity classification |
| Drop-in for existing Claude integrations | Some edge cases still need Sonnet |
| Transparent routing with full audit trail | Extra latency (~50ms) for classification |
Verdict
If you're running Claude at scale and not routing by complexity, you're burning money. save-money is the obvious first skill to install on mr.technology. The ROI is immediate — if Haiku handles even 30% of your volume, you're cutting the bill significantly.