Fully hosted web agents

Browser Use Agents

Describe the task in a sentence. An agent runs it in a real stealth browser and returns structured data.

A hand holding a magnifier over a train of glass carriages, each carrying a city

Measured on real browser tasks.

Compare task success and cost on the same benchmark. Results below use our 106-task Internal Bench Hard set; the benchmark page explains the scoring and cost calculation.

Browser Use: 82% of tasks solved at 17¢ each. That is 20 points better than Opus 5, which costs 20× more per solved task.

Toolstrict accuracyCost

Browser Use82%$0.17

Opus 562%$3.40

Gemini 3.1 Pro59%$2.20

Sonnet 559%$1.55

GPT-5.652%$1.10

Gemini 3.6 Flash46%$0.62

GPT-537%$0.44

Internal Bench Hard · updated 2026-08-01 · all benchmarks

Every agent benchmark

Thousands of agents in parallel

Every run gets its own isolated session: a stealth browser, a proxy exit and a model context. Fan out a queue of work and collect the results.

One cell = one concurrent run · each with its own browser, proxy and model context

Runs on our own stealth browsers

The agent drives the same infrastructure behind stealth browsers, so it reaches sites that block a stock headless browser on the first request.

Browser Use81%

Stock headless Chrome2%

Protected sites reached on our stealth benchmark

Real-time web monitoring

Describe what to watch in a sentence and get told when it changes. No scrapers to maintain, no selectors to repair, no diffing logic to write.

One tick = one check · the tall ones are the runs that came back different

Structured outputs

Results come back as JSON you can hand straight to the next step, with webhook callbacks when a run finishes.

{
  "title": "Aeron Chair",
  "price": 1395.00,
  "currency": "USD",
  "in_stock": true
}

You define the shape · the agent fills it in

One sentence in, structured data out.

There is no flow to build and no selectors to maintain. Post the task, pick a model, and the agent handles the clicking, scrolling, logins and retries on the other side.

Read the quickstart

# Hand the agent a task in plain Englishcurl https://api.browser-use.com/api/v4/runs \  -H "X-Browser-Use-API-Key: $BROWSER_USE_API_KEY" \  -H "Content-Type: application/json" \  -d '{    "task": "Find the top Hacker News story and return its title and points"  }' # {#   "id": "9c2f...",#   "status": "queued",#   "sessionId": "5b81...",#   "eventsUrl": "https://api.browser-use.com/..."# }

run-agent.sh·Pick any model · swap it per run

Give your product an agent.

Run your first agent

FAQ

A web agent is an LLM that drives a real browser: it reads the page, decides what to click or type, and works through a task the way a person would. That is the difference from a scraper, which follows selectors someone wrote by hand and breaks the moment the page changes.