Cost and depth
npx monkeytest run https://example.com --depth basicimport { resolveDepthPreset, BASIC_PRESET } from '@monkeytestai/monkeytester';
const quick = resolveDepthPreset('basic');
// quick.label === 'Quick pass'
// quick.maxPages === 6
// quick.maxFlows === 5
// quick.focus === 'forms-search-cta'
// BASIC_PRESET.totalBudgetMs === 120_000Cost follows the page cap, the flow cap, and which model you assign to each phase. Page checks and auth-negative checks do not call the planner. They still open the browser.
Presets
| Depth | Id | Pages | LLM flows | Also |
|---|---|---|---|---|
| Quick pass | basic | 6 | 5 | Focus forms-search-cta. Up to 5 page checks and 6 auth-negative flows, outside the LLM budget. 10s step, 25s flow, 12s page, 120s total. |
| Shallow | shallow | 4 | 3 | Smaller smoke preset |
| Broader pass | standard | 8 | 6 | The CLI default caps, when you do not pass --depth |
| Deeper dig | deep | 16 | 10 |
--max and --flows override the page and flow caps. Aliases: quick-pass, broader-pass, deeper-dig.
Quick pass is about 10 cases and usually finishes in 1–2 minutes. The 120 second budget is a hard cap: past it, remaining page checks are skipped instead of failing the run.
Estimates for a standard run
The figures below are order-of-magnitude estimates for one standard run (8 pages, 6 flows, about 6 steps each), using these token assumptions per call:
| Phase | Input | Output |
|---|---|---|
| Explorer, per page | ~2,500 | ~250 |
| Planner, once | ~6,000 | ~1,500 |
| Runner, per step | ~1,500 | ~120 |
That is about 80k input tokens and 8k output tokens. A shallow run is smaller. A deep run is larger. Treat the dollars as a sketch, not an invoice. Provider prices move.
| Provider | Cheap everywhere | Balanced | Smart everywhere |
|---|---|---|---|
| OpenAI | gpt-4o-mini ~$0.02 | gpt-4o-mini plus gpt-4o planner ~$0.05 | gpt-4o ~$0.45 |
| Gemini | gemini-flash-lite-latest ~$0.01 | flash-lite plus gemini-flash-latest planner ~$0.02 | gemini-flash-latest ~$0.05 |
| Anthropic | claude-haiku-4-5 ~$0.08 | Haiku plus claude-sonnet-4-5 planner ~$0.18 | claude-sonnet-4-5 ~$0.90 |
| Mistral | mistral-small-latest ~$0.03 | small plus mistral-large-latest planner ~$0.06 | mistral-large-latest ~$0.18 |
| Groq | llama-3.3-70b-versatile ~$0.06 | same model ~$0.06 | same model ~$0.06 |
| Local (Ollama) | electricity | electricity | electricity |
Prompt caching on OpenAI and Anthropic can cut a repeated planner prefix. See Prompt caching. The cheapest lever is a smaller --max, then a smaller --flows, then a smart planner with a cheap runner.
MODEL_PRICES and estimateCostUsd are exported if you want to price a model id in your own host. They are not a bill.