Cost management
How much Nitpick costs, what drives cost, and how to control it.
Nitpick is free. Your LLM provider isn’t. Here’s what to expect.
Pricing baseline
Check anthropic.com/pricing for current rates. Approximate:
| Model | Input | Output |
|---|---|---|
| claude-opus-4-6 | ~$15/M | ~$75/M |
| claude-sonnet-4-6 | ~$3/M | ~$15/M |
| claude-haiku-4-5 | ~$1/M | ~$5/M |
Anthropic’s prompt caching charges ~10% of input cost for cached tokens. Nitpick enables caching by default for the skill’s system prompt (~15-20k tokens), which is re-read on every agent iteration.
Per-component cost
| Component | LLM calls? | Rough cost |
|---|---|---|
nitpick crawl | No (deterministic Playwright) | $0 |
nitpick run --scope smoke | No | $0 |
| Scope inference (iterative) | 1 call, fast model | ~$0.01 |
| One simple page (static dashboard) | ~20 calls | $1–2 (with caching) |
| One complex page (8-field form) | ~30 calls | $2–4 (with caching) |
| One wizard step | ~25 calls | $1.50–3 (with caching) |
| Flow test (3 stages) | ~75 calls | $6–12 (with caching) |
Run-level estimates
| Scope | Size | Cost (Opus + caching) | Cost (Sonnet + caching) |
|---|---|---|---|
smoke | 15 pages | $0 | $0 |
targeted | 1 page | $1–4 | $0.30–1 |
iterative | 3 pages + 1 flow | $10–20 | $2–5 |
full | 15 pages + 1 flow | $30–50 | $8–15 |
Cost levers
- Switch to Sonnet: ~5x cheaper than Opus for most tasks.
- Use Haiku for fast_model: scope inference uses this by default.
- Cap
max_iterations: a misbehaving agent can spiral. Set to 30-50 for CI. - Scope tight: iterative with a specific change description tests only affected pages.
- Skip flows unless changed: flows are the most expensive.
- Use smoke frequently: free, no LLM.
What caching saves
Before caching: ~$8–12 per page on Opus. After caching: ~$1.50–3 per page on Opus.
The skill’s system prompt (~15-20k tokens) is read on every agent iteration. Without caching, you pay full input cost for it 20+ times per page. With caching, ~90% discount on iterations 2+.
Cache TTL is 5 minutes. Serial iterations within one page hit the cache. Parallel page tests each get their own cache (not shared).
Budget caps
A future release will add llm.max_dollars_per_run. Until then:
- Set
max_iterationsconservatively (30-50) - Start with Sonnet for new apps
- Monitor your Anthropic console dashboard
- Set account-level spend limits in the provider’s console
Cost vs hiring
A US senior QA engineer costs ~$120-150k/year. Nitpick running a full regression every night with Opus costs ~$1,000-1,500/month ($30/night × 30 days). The delta isn’t the LLM bill it’s that Nitpick runs in parallel, overnight, and doesn’t quit.