Model Tuner
Pay less for the same quality
npx promptreports-cli modelsCategorizes every turn by task type (search, read, write, complex) and suggests model downgrades. Opus is for architecture. Not grep.
models is the router optimizer. It classifies every turn by task type — is this a simple grep? a multi-file refactor? a high-level architecture decision? Then it checks which model you used. If you used Opus ($15/M in, $75/M out) for 30% simple searches, it suggests Sonnet or Haiku for those — with confidence ratings. Shows dollar savings per week and month.
On this page
Prerequisites
- Session files with turn data
Flags & Options
| Flag | Description | Default |
|---|---|---|
| --days N | Lookback | 7 |
| --optimize | Show optimization suggestions (default on) | — |
| --compare | Compare model pricing table | — |
| --savings | Calculate potential savings | — |
| --json | JSON output | — |
Examples
Default (with suggestions)
npx promptreports-cli modelsWhat can you downgrade?
30-day savings
npx promptreports-cli models --savings --days 30Monthly optimization potential.
Price comparison
npx promptreports-cli models --compareSee Opus vs Sonnet vs Haiku pricing side-by-side.
Output
Model usage breakdown + task distribution + prioritized downgrade suggestions with confidence + estimated savings.
Model Usage
Model Calls Tokens Cost %
claude-opus 18 234.5K $8.45 78%
claude-sonnet 5 42.3K $1.20 14%
claude-haiku 2 8.1K $0.12 2%
Task Distribution
claude-opus: search 30%, read 15%, write 35%, complex 20%
Optimization Suggestions
⚡ 30% of claude-opus calls (search tasks) → claude-sonnet
Search/grep tasks — sonnet handles equally well
Confidence: high | Savings: ~$1.45/week
⚡ 15% of claude-opus calls (read tasks) → claude-haiku
Simple reads rarely need Opus
Confidence: medium | Savings: ~$0.68/week
──────────────────────────────
Total estimated savings: $2.13/week ($9.13/month)What it reads and writes
Reads
- ~/.claude/projects/**/*.jsonl
Writes
Nothing (read-only)
Free vs Pro usage
Free tier
- See where you're overpaying
- Get specific downgrade suggestions with dollar impact
- Apply manually by changing your default model
Pro tier
Upgrade- Auto-routing: CLI injects model preferences into your Claude Code session based on task type
- Team-wide routing policies
- Quality tracking — ensure downgrades don't hurt output quality
- Per-user routing (senior devs get Opus, CI bots get Haiku)
Pro tips
- Start with high-confidence suggestions — lowest risk
- Haiku 4.5 is surprisingly capable for read/summarize tasks
- Opus is for architecture, security, and hard debugging — not formatting