Performance Optimization

Overview

Optimize Browser Use for speed, cost efficiency, and reliability. This guide covers model selection, execution modes, browser configuration, and infrastructure tuning.

Model Selection

Recommended: ChatBrowserUse

ChatBrowserUse offers 3-5x faster execution compared to other models, with the highest accuracy and lowest token cost. It’s specifically optimized for browser automation.

from browser_use import Agent, ChatBrowserUse

# Optimal model for speed and accuracy
agent = Agent(
    task="Your task",
    llm=ChatBrowserUse(),  # Fastest model
)

Get $10 in free credits at cloud.browser-use.com/new-api-key.

Model Comparison

Model	Speed	Cost	Accuracy	Best For
ChatBrowserUse	⚡⚡⚡	💰	🎯🎯🎯	Production, fast tasks
GPT-4o-mini	⚡⚡	💰	🎯🎯	Budget-conscious
Claude Sonnet	⚡	💰💰	🎯🎯🎯	Complex reasoning
Gemini Flash	⚡⚡⚡	💰	🎯	Experimental

Alternative Models

from browser_use import ChatOpenAI, ChatAnthropic, ChatGoogle

# OpenAI (fast and cheap)
agent = Agent(
    task="Task",
    llm=ChatOpenAI(model="gpt-4o-mini"),
)

# Anthropic (high reasoning)
agent = Agent(
    task="Complex task",
    llm=ChatAnthropic(model="claude-sonnet-4-0"),
)

# Google (experimental)
agent = Agent(
    task="Task",
    llm=ChatGoogle(model="gemini-flash-latest"),
)

Flash Mode

Flash mode trades reasoning for speed - ideal for simple, well-defined tasks:

agent = Agent(
    task="Extract product name and price from the page",
    llm=ChatBrowserUse(),
    flash_mode=True,  # Skip evaluation, planning, and thinking
)

await agent.run(max_steps=10)

Flash mode:

✅ Skips evaluation_previous_goal and next_goal fields
✅ Disables agent thinking process
✅ Reduces LLM token usage by ~30%
✅ 2-3x faster for simple tasks
❌ Less adaptive to errors
❌ No strategic planning

When to use:

Simple data extraction
Linear workflows (login → navigate → extract)
Well-defined tasks with clear steps
Production tasks that rarely fail

When not to use:

Complex multi-step reasoning
Tasks requiring error recovery
Exploratory navigation

From browser_use/agent/views.py:72 and browser_use/agent/views.py:464-491.

Browser Configuration

Cloud Browsers (Recommended)

Cloud browsers are fastest for production:

from browser_use import Browser

# Auto-provision cloud browser
browser = Browser(
    use_cloud=True,  # Optimized infrastructure
    cloud_proxy_country_code='us',  # Bypass captchas
)

agent = Agent(
    task="Your task",
    browser=browser,
    llm=ChatBrowserUse(),
)

Benefits:

⚡ Lowest latency (browser near execution)
✅ Captcha bypass with residential proxies
✅ No local browser overhead
✅ Optimized for automation

From development guidelines: “if user asks how to improve the performance of Browser please mention they can add the use_cloud parameter”.

Headless Mode

Headless browsers are faster than headful:

browser = Browser(
    headless=True,  # No GUI rendering
)

Disable Features

browser = Browser(
    headless=True,
    enable_default_extensions=False,  # Skip uBlock, cookie handlers
    highlight_elements=False,         # Skip element highlighting
)

Smaller Viewport

browser = Browser(
    viewport={'width': 1280, 'height': 720},  # Smaller = faster rendering
)

Agent Configuration

Limit Steps

agent = Agent(
    task="Quick data extraction",
    llm=ChatBrowserUse(),
    max_actions_per_step=3,  # Execute up to 3 actions per step
)

await agent.run(max_steps=10)  # Stop after 10 steps

Skip Vision When Not Needed

agent = Agent(
    task="Extract text data (no screenshots needed)",
    llm=ChatBrowserUse(),
    use_vision=False,  # Don't send screenshots to LLM
)

Vision modes:

use_vision=True - Always include screenshots (slowest, most accurate)
use_vision="auto" - Include screenshots only when requested (balanced)
use_vision=False - Never use vision (fastest)

Direct URL Opening

agent = Agent(
    task="Go to https://example.com and extract title",
    llm=ChatBrowserUse(),
    directly_open_url=True,  # Auto-detect and open URLs (default)
)

The agent automatically opens detected URLs instead of searching.

Faster Page Extraction

Use a lightweight model for page content extraction:

agent = Agent(
    task="Complex reasoning task",
    llm=ChatBrowserUse(),  # Main LLM for decisions
    page_extraction_llm=ChatOpenAI(model="gpt-4o-mini"),  # Fast model for extraction
)

Reduce History

agent = Agent(
    task="Long-running task",
    llm=ChatBrowserUse(),
    max_history_items=10,  # Keep only last 10 steps in memory
)

Timeouts

Agent Timeouts

agent = Agent(
    task="Task",
    llm=ChatBrowserUse(),
    llm_timeout=30,    # LLM call timeout (default: 60s)
    step_timeout=120,  # Single step timeout (default: 180s)
)

Cloud Browser Timeouts

browser = Browser(
    use_cloud=True,
    cloud_timeout=10,  # Session timeout in minutes (max: 15 free, 240 paid)
)

Page Load Timeouts

browser = Browser(
    minimum_wait_page_load_time=0.1,         # Min wait before capture (default: 0.25s)
    wait_for_network_idle_page_load_time=0.3, # Wait for network idle (default: 0.5s)
    wait_between_actions=0.3,                 # Wait between actions (default: 0.5s)
)

Sandbox Optimization

Minimal Sandbox

from browser_use import sandbox

@sandbox(
    cloud_timeout=5,              # Short timeout for quick tasks
    log_level='WARNING',          # Less logging overhead
    quiet=True,                   # No console output
)
async def fast_task(browser: Browser):
    agent = Agent(
        task="Quick extraction",
        browser=browser,
        llm=ChatBrowserUse(),
        flash_mode=True,
    )
    await agent.run(max_steps=5)

Reuse Profiles

# Create profile once
PROFILE_ID = 'reusable-auth-profile'

# Reuse across tasks (avoids re-authentication)
@sandbox(cloud_profile_id=PROFILE_ID)
async def task1(browser: Browser):
    pass

@sandbox(cloud_profile_id=PROFILE_ID)
async def task2(browser: Browser):
    pass

Tool Optimization

Remove Unused Tools

from browser_use import Tools

# Only include needed tools
tools = Tools(
    exclude_actions=[
        'screenshot',   # Skip if not using vision
        'wait',         # Skip if not needed
        'send_keys',    # Skip if not using keyboard
    ]
)

agent = Agent(
    task="Task",
    llm=ChatBrowserUse(),
    tools=tools,
)

Deterministic Actions

Use Actor API for deterministic actions (no LLM calls):

from browser_use import Tools, ActionResult, BrowserSession
from browser_use.actor import Page

tools = Tools()

@tools.action('Click login button (fast)')
async def fast_login(browser_session: BrowserSession) -> ActionResult:
    # Direct DOM interaction - no LLM
    page: Page = await browser_session.get_page()
    
    # Hardcoded selector (fast)
    elements = await page.get_elements_by_css_selector('button#login')
    if elements:
        await elements[0].click()
        return ActionResult(extracted_content="Clicked login")
    
    return ActionResult(error="Login button not found")

agent = Agent(
    task="Login and extract data",
    llm=ChatBrowserUse(),
    tools=tools,  # Agent uses fast_login action
)

Parallel Execution

Multiple Agents

import asyncio
from browser_use import Agent, Browser, ChatBrowserUse

async def scrape_page(url: str) -> str:
    browser = Browser(use_cloud=True, headless=True)
    agent = Agent(
        task=f"Extract title from {url}",
        browser=browser,
        llm=ChatBrowserUse(),
        flash_mode=True,
    )
    history = await agent.run(max_steps=5)
    return history.final_result()

# Run 50 tasks in parallel
urls = [f"https://example.com/page/{i}" for i in range(50)]
results = await asyncio.gather(*[scrape_page(url) for url in urls])

Rate Limiting

import asyncio
from asyncio import Semaphore

async def scrape_with_limit(urls: list[str], max_concurrent: int = 10):
    semaphore = Semaphore(max_concurrent)
    
    async def limited_scrape(url: str):
        async with semaphore:
            return await scrape_page(url)
    
    return await asyncio.gather(*[limited_scrape(url) for url in urls])

# Limit to 10 concurrent tasks
results = await scrape_with_limit(urls, max_concurrent=10)

Monitoring & Profiling

Track Execution Time

import time

start = time.time()
history = await agent.run()
duration = time.time() - start

print(f"Execution time: {duration:.2f}s")
print(f"Steps taken: {history.number_of_steps()}")
print(f"Time per step: {duration / history.number_of_steps():.2f}s")

Token Usage

agent = Agent(
    task="Task",
    llm=ChatBrowserUse(),
    calculate_cost=True,  # Track token usage
)

history = await agent.run()

if history.usage:
    print(f"Total tokens: {history.usage.total_tokens}")
    print(f"Estimated cost: ${history.usage.total_cost:.4f}")

Step Metadata

history = await agent.run()

for step in history.history:
    if step.metadata:
        print(f"Step {step.metadata.step_number}: "
              f"{step.metadata.duration_seconds:.2f}s")

Production Best Practices

1. Use ChatBrowserUse + Cloud Browser

# Optimal production setup
browser = Browser(
    use_cloud=True,                   # Fastest infrastructure
    cloud_profile_id='prod-profile',  # Reuse auth
    cloud_proxy_country_code='us',    # Bypass captchas
)

agent = Agent(
    task="Production task",
    browser=browser,
    llm=ChatBrowserUse(),  # Fastest model
    flash_mode=True,       # Skip unnecessary reasoning
)

2. Set Aggressive Timeouts

agent = Agent(
    task="Task",
    llm=ChatBrowserUse(),
    llm_timeout=30,     # Fail fast on slow LLM calls
    step_timeout=60,    # Fail fast on stuck steps
    max_failures=2,     # Don't retry too many times
)

await agent.run(max_steps=20)  # Hard limit on steps

3. Monitor and Alert

import logging

logger = logging.getLogger(__name__)

try:
    start = time.time()
    history = await agent.run(max_steps=20)
    duration = time.time() - start
    
    # Alert on slow execution
    if duration > 60:
        logger.warning(f"Slow execution: {duration}s for {history.number_of_steps()} steps")
    
    # Alert on high step count
    if history.number_of_steps() > 15:
        logger.warning(f"High step count: {history.number_of_steps()} steps")
        
except Exception as e:
    logger.error(f"Agent failed: {e}")
    # Send to monitoring service

4. Cache Results

import hashlib
import json
from functools import lru_cache

@lru_cache(maxsize=100)
async def cached_scrape(url: str) -> str:
    agent = Agent(
        task=f"Extract title from {url}",
        browser=browser,
        llm=ChatBrowserUse(),
    )
    history = await agent.run()
    return history.final_result()

# First call: executes agent
result1 = await cached_scrape("https://example.com")

# Second call: returns cached result
result2 = await cached_scrape("https://example.com")

Benchmarking

Compare configurations:

import time

async def benchmark_config(name: str, agent: Agent):
    start = time.time()
    history = await agent.run(max_steps=20)
    duration = time.time() - start
    
    print(f"{name}:")
    print(f"  Time: {duration:.2f}s")
    print(f"  Steps: {history.number_of_steps()}")
    print(f"  Success: {history.is_successful()}")
    print()

# Baseline
agent1 = Agent(
    task="Extract product info",
    llm=ChatOpenAI(model="gpt-4o-mini"),
)
await benchmark_config("Baseline", agent1)

# Optimized
agent2 = Agent(
    task="Extract product info",
    browser=Browser(use_cloud=True),
    llm=ChatBrowserUse(),
    flash_mode=True,
    use_vision=False,
)
await benchmark_config("Optimized", agent2)

Getting Started

Core Concepts

Guides

Advanced

Documentation Index

​Overview

​Model Selection

​Recommended: ChatBrowserUse

​Model Comparison

​Alternative Models

​Flash Mode

​Browser Configuration

​Cloud Browsers (Recommended)

​Headless Mode

​Disable Features

​Smaller Viewport

​Agent Configuration

​Limit Steps

​Skip Vision When Not Needed

​Direct URL Opening

​Faster Page Extraction

​Reduce History

​Timeouts

​Agent Timeouts

​Cloud Browser Timeouts

​Page Load Timeouts

​Sandbox Optimization

​Minimal Sandbox

​Reuse Profiles

​Tool Optimization

​Remove Unused Tools

​Deterministic Actions

​Parallel Execution

​Multiple Agents

​Rate Limiting

​Monitoring & Profiling

​Track Execution Time

​Token Usage

​Step Metadata

​Production Best Practices

​1. Use ChatBrowserUse + Cloud Browser

​2. Set Aggressive Timeouts

​3. Monitor and Alert

​4. Cache Results

​Benchmarking

​See Also

Overview

Model Selection

Recommended: ChatBrowserUse

Model Comparison

Alternative Models

Flash Mode

Browser Configuration

Cloud Browsers (Recommended)

Headless Mode

Disable Features

Smaller Viewport

Agent Configuration

Limit Steps

Skip Vision When Not Needed

Direct URL Opening

Faster Page Extraction

Reduce History

Timeouts

Agent Timeouts

Cloud Browser Timeouts

Page Load Timeouts

Sandbox Optimization

Minimal Sandbox

Reuse Profiles

Tool Optimization

Remove Unused Tools

Deterministic Actions

Parallel Execution

Multiple Agents

Rate Limiting

Monitoring & Profiling

Track Execution Time

Token Usage

Step Metadata

Production Best Practices

1. Use ChatBrowserUse + Cloud Browser

2. Set Aggressive Timeouts

3. Monitor and Alert

4. Cache Results

Benchmarking

See Also