Claude AI Agent Scraped Google Maps Accurately but Used 190K–540K Tokens Per Place

A developer tested a Claude Sonnet AI agent equipped only with browser tools to scrape data from 60 Google Maps coffee-shop listings across New York, San Francisco, and Chicago, without writing any pre-built scraper code. The agent successfully collected name, rating, review count, address, and phone number for every listing, achieving full accuracy across all four test runs. However, each place took 23–35 seconds and consumed 190,000–540,000 processed tokens, compared to just 6–14 seconds and zero model tokens for an equivalent Playwright script. Costs were not flat — a longer Chicago run of 30 places turned out cheaper per record after the agent spontaneously switched to a more efficient JavaScript strategy targeting aria-label attributes. The experiment highlights that while LLM agents can replace scrapers in terms of accuracy, they carry significant and variable token overhead that grows as context accumulates across a session.