Cost & depth

How much does a MonkeyTest run cost? It depends on:

1. Depth — how many pages you crawl (maxPages) and how many flows you run (maxFlows × steps).
2. Models — which LLM you give to each phase.
3. Page size — bigger pages = bigger explorer prompts.

This page gives you estimated costs for the three depth presets (shallow, standard, deep) across all supported providers, plus the assumptions and a calculator so you can plug in your own numbers.

Token prices change every month

We update this page periodically, but always sanity-check against the provider's pricing page before you budget. Treat these numbers as order of magnitude, not invoices.

Depth presets

Depth	maxPages	maxFlows	Avg steps/flow	Total LLM calls
Shallow (smoke)	4	3	4	~16 (4 explore + 1 plan + 12 run)
Standard (default)	8	6	6	~45 (8 + 1 + 36)
Deep (release)	16	10	8	~97 (16 + 1 + 80)

Token assumptions

Conservative averages from the engine's actual prompts (you can see them all under monkeytester/src/{explorer,planner,runner,triager}/).

Phase	Input tokens / call	Output tokens / call
Explorer (per page)	~2,500	~250
Planner (one shot)	~6,000	~1,500
Runner (per step)	~1,500	~120

So a standard run uses roughly:

• Explorer: 8 × (2,500 in + 250 out) = 20,000 in / 2,000 out
• Planner: 1 × (6,000 in + 1,500 out) = 6,000 in / 1,500 out
• Runner: 36 × (1,500 in + 120 out) = 54,000 in / 4,320 out
• Total ≈ 80k input tokens, 8k output tokens

A shallow run is ~25k in / 2k out. A deep run is ~175k in / 17k out.

Per-provider cost estimates

USD per standard run (8 pages, 6 flows, ~6 steps each). Calculated against published list prices as of writing.

Provider	Cheap-everywhere	Balanced (cheap explorer/runner + smart planner)	Smart-everywhere
OpenAI	gpt-4o-mini everywhere → ~$0.02	gpt-4o-mini + gpt-4o planner → ~$0.05	gpt-4o everywhere → ~$0.45
Gemini	flash-lite-latest everywhere → ~$0.01	flash-lite-latest + flash-latest planner → ~$0.02	flash-latest everywhere → ~$0.05
Anthropic	claude-haiku-4-5 everywhere → ~$0.08	haiku + sonnet-4-5 planner → ~$0.18	claude-sonnet-4-5 everywhere → ~$0.90
Mistral	mistral-small-latest everywhere → ~$0.03	small + large-latest planner → ~$0.06	mistral-large-latest everywhere → ~$0.18
Groq (Llama 3.3 70B)	flat → ~$0.06	(no separate smart tier) → ~$0.06	flat → ~$0.06
Local (Ollama)	electricity only	electricity only	electricity only

For shallow runs, divide by ~3. For deep runs, multiply by ~2.5. The smart planner alone accounts for most of the variance — that's the lever to pull when budget matters.

These are list prices

Volume discounts, batch APIs, prompt-caching (OpenAI / Anthropic) and free tiers can dramatically reduce real costs. Anthropic's prompt caching is especially impactful when you call the planner many times against the same site map.

Cost calculator

For your own combinations:

cost ≈
    (explorer_calls × explorer_in × P_in_explorer + explorer_calls × explorer_out × P_out_explorer)
  + (1             × planner_in  × P_in_planner  + 1              × planner_out  × P_out_planner)
  + (runner_calls  × runner_in   × P_in_runner   + runner_calls   × runner_out   × P_out_runner)

Where:

• explorer_calls ≈ maxPages
• runner_calls ≈ maxFlows × avg_steps_per_flow
• P_in_*, P_out_* are the per-token prices for the model you assigned to that phase.

Example (standard run, mixed setup): Gemini Flash Lite explorer + Claude Sonnet 4.5 planner + Groq Llama runner.

explorer:  8 × (2500 × $0.075 + 250 × $0.30)  / 1e6  ≈ $0.0021
planner:   1 × (6000 × $3.00  + 1500 × $15.00) / 1e6 ≈ $0.041
runner:   36 × (1500 × $0.59  + 120  × $0.79)  / 1e6 ≈ $0.035
                                                       ─────────
                                                       ≈ $0.078

Cost-control tips

1. Lower maxPages first. Halving pages roughly halves explorer cost, and the planner usually has enough info from 4–6 pages.
2. Lower maxFlows second. This linearly scales runner cost — the largest line item in deep runs.
3. Use a smart planner, cheap everything else. This is the single best price/quality trade-off.
4. Use Groq for the runner. Hot path, near-zero latency, flat low pricing.
5. Use prompt caching where supported (OpenAI gpt-4o, Anthropic). The site map changes per URL; the system prompts don't.
6. Cap pageTimeoutMs if the explorer is hitting slow pages — wasted seconds = wasted explorer LLM calls when retries fire.
7. Run locally on Ollama for development. Quality drops, but cost goes to zero.

A pragmatic default

For a small team running on every PR:

import { Monkey, autoLLMConfig } from '@monkeytestai/monkeytester';

const monkey = new Monkey({
  llm: autoLLMConfig({
    quality: 'balanced',
    overrides: { runner: { provider: 'groq', model: 'llama-3.3-70b-versatile' } },
  }),
});

await monkey.executeAndTriage(url, {
  explore: { maxPages: 6, screenshotsDir },
  plan:    { maxFlows: 4 },
  run:     { screenshotsDir, headless: true },
});

Estimated cost per PR: ~$0.04 with OpenAI keys, or ~$0.02 with Gemini.

Next: Live events →