Measured, not claimed.
Live tasks on real sites, run on pilotfish and on Browser Use with the same model and country. Every figure here comes from those runs or from a published leaderboard.
The app, against Browser Use Cloud
Shopping, travel, property and finance tasks, each run on both services with GPT-6 Luna and a US browser.
On average, against Browser Use Cloud.
On average, residential traffic included.
On average, on the same model.
The open MCP, against Browser Use's
Lookups driven by the same coding agent and model (Sonnet 5.5 in Claude Code), once through each free MCP.
On average, against Browser Harness.
On average, on the same model.
On average, against Browser Harness.
On the leaderboards
Public benchmarks of live web tasks, each scored as its leaderboard publishes it. pilotfish is placed by its own runs against the leader on the same tasks.
WebVoyager
Tasks done, as published- 1browser-control (Fable 5)self-reported99.2%
- 2Alumnium98.5%
- 3Surfer 297.1%
- 4Magnitude93.9%
- 5Surfer-H + Holo192.2%
- 6pilotfish92.0%
- 7Browserable90.4%
- 8Browser Use89.1%
- 9OpenAI Operator87.0%
- 10Skyvern 2.085.9%
- 11Google Project Mariner83.5%
26% fewer failed tasks than browser-use on the same tasks and judge, applied to Browser Use's published 89.1%.
Other rows: Steel's WebVoyager leaderboard, October 2026.
Online-Mind2Web
Tasks done of 300, as published- 1Browser Use Cloud97.0%
- 2pilotfish96.7%
- 3GPT-5.4 computer usejudge unpublished93.0%
- 4UI-TARS-288.0%
- 5ABP + Claude Opus 4.686.0%
- 6OpenAGI Lux84.0%
- 7TinyFish81.0%
- 8Yutori Navigator79.0%
- 9ChatGPT Atlas agent71.0%
- 10Google Gemini CUA69.0%
9% more failed tasks than browser-use on the same tasks and judge, applied to Browser Use Cloud's published 97%.
Other rows: Browser Use's Online-Mind2Web results, March 2026.
Shopping agents
Stores done, of 500- 1pilotfishTied 1st
- 1Browser Use Cloud91.8%
- 3Browser Use On-Prem89.0%
- 4Grok Botclaimed76.2%
- 5Dot72.6%
- 6Meta Museself-graded61.2%
Tied with Browser Use Cloud, the leader, on our head-to-head tasks.
Other rows: Browser Use's published results.
Browser tools for coding agents
Tasks done per 10M tokens- 1pilotfish open MCP1st
- 2Browser Harness4.7
- 3agent-browser3.0
- 4Playwright MCP2.3
- 5Chrome DevTools MCP1.5
As many tasks as Browser Harness, the leader, on 18% fewer tokens.
Other rows: Browser Use's published results.
Why it costs less
Most of an agent's bill is the model reading pages. pilotfish gives it less to read, and less to wait for.
Each screen sent once
After an action, the model gets only what changed on the page, not the whole page again. Fewer tokens on every step.
Compact pages
Pages reach the model as short outlines of what's on screen, with the markup and repetition stripped out.
Planning while the browser opens
The first plan is made during the seconds the browser takes to start, so the work begins the moment it's ready.
Smaller paid traffic
Scripts and fonts that sites share are fetched once and cached, so the residential IP carries pages, not the same files again.
How we test
The rules every figure on this page follows.
- 01
Live sites, same days
Real shopping, travel, property and finance tasks, run on both products within the same days, on the live sites.
- 02
Same model, same country
Both sides use the same AI model and browse from the same country, so only the agent differs.
- 03
Billed cost, real time
Cost is what each service billed for the task, residential traffic included. Time runs from sending the task to the answer.
- 04
Right answers only
A task counts as done only when its final answer was correct. Public test sets are scored by each benchmark's own judge.
Hand off your first task today.
Create your account with Google in a minute, add credits, and hand it something tedious.