Anthropic Computer Use vs Google Project Mariner: Browser Agents Comparison
Compare Anthropic Computer Use and Google Project Mariner side by side. Both are tools in the Browser Agents category.
Updated March 1, 2026
The short answer
Choose Anthropic Computer Use if generalizes to unknown software without pre-programming.
Choose Google Project Mariner if multimodal understanding allows parsing of text, images, buttons, forms, and code on web pages.
Quick Comparison
| Category | Browser Agents | Browser Agents |
| Website | docs.anthropic.com | deepmind.google |
Pros & Cons: Anthropic Computer Use vs Google Project Mariner
- +Generalizes to unknown software without pre-programming
- +Available through Anthropic API, AWS Bedrock, and Google Cloud Vertex AI
- +Zoom capability for fine-grained interaction with dense interfaces
- +MCP integration enables secure access to local data and tools
- +Foundational capability powering the browser agent ecosystem
- −Security risks from prompt injection when interacting with the internet
- −Notably slower than human-directed computer actions
- −Hallucinated clicks and coordinate errors require safety guardrails
- −High token consumption from iterative screenshot loops
- +Multimodal understanding allows parsing of text, images, buttons, forms, and code on web pages
- +Teach & Repeat feature enables learning custom workflows by demonstration
- +Can execute up to 10 parallel tasks for comparing prices or gathering data from multiple sources
- +Powered by Gemini 2.0, providing advanced AI capabilities and reasoning
- +Integration with Google's ecosystem provides access to additional tools and features
- −Premium pricing at $249.99/month is significantly higher than most consumer AI subscriptions
- −Currently limited to US users, though expanding to more countries
- −Still experimental with slow performance and prone to roadblocks like CAPTCHA or login issues
- −Cannot fill out credit card numbers, billing information, accept cookies, or sign terms of service
- −Requires significant computational resources with virtual machine architecture
Pricing: Anthropic Computer Use vs Google Project Mariner
- · Most accurate for computer use
- · Screenshot interpretation
- · Keyboard and mouse simulation
- · 466-499 extra system prompt tokens
- · Most capable model overall
- · Computer use support
- · Complex reasoning tasks
- · Reduce repeated context costs
- · Available for all computer use models
- · Access to Project Mariner browser agent
- · Advanced Gemini 2.0 models and features
- · Up to 10 concurrent browser tasks
- · Teach & Repeat workflow learning
How to Choose a Browser Agents Tool
Key criteria to evaluate when comparing Browser Agents solutions:
About Anthropic Computer Use
Anthropic Computer Use is Claude's native ability to interact with computer interfaces by clicking, typing, scrolling, and navigating desktop and browser environments. First introduced in October 2024 with Claude 3.5 Sonnet, it has been expanded to Claude Sonnet 4.5, Sonnet 4.6, and Opus 4.6.
Computer Use enables Claude to autonomously operate any GUI application, making it the foundational model capability that tools like Browser Use, Stagehand, and Browserbase build upon. The feature includes a Zoom capability for high-resolution screen region analysis and integrates with the Model Context Protocol (MCP) for secure local data access.
Anthropic was founded in 2021 by Dario and Daniela Amodei and is headquartered in San Francisco. The company has grown to approximately 4,074 employees, reached a $380B valuation with a $30B Series G round in February 2026, and generates approximately $14B in annualized revenue.
About Google Project Mariner
Project Mariner is Google DeepMind's experimental AI browser agent that uses Gemini's powerful multimodal capabilities to autonomously navigate websites, understand screen content, plan tasks, and execute them by clicking, typing, scrolling, and filling forms. Powered by Gemini 2.0, Mariner represents a significant advancement in human-agent interaction, starting with browsers as the primary interface. The agent can parse text, images, buttons, forms, and code on web pages, allowing it to navigate complex sites much like a human would.
Key capabilities include the Observe-Plan-Act loop for intelligent task execution, Teach & Repeat functionality where users can demonstrate a workflow once and Mariner will learn to replicate it in future runs, and multimodal understanding that enables the agent to comprehend various types of web content. Thanks to its virtual machine architecture, Mariner can execute up to 10 tasks concurrently, making it useful for comparing prices across sites or gathering data from multiple sources simultaneously. Practical applications include using resume information to find personalized job listings, researching topics to summarize information from across the web, planning travel, and discovering local spots.
As of May 2025, Project Mariner moved from experimental prototype to integrated ecosystem feature, becoming available to Google AI Ultra subscribers in the US. Google DeepMind, the parent organization, was founded in November 2010 and is headquartered in Kings Cross, London. Following a merger with Google AI's Google Brain division in April 2023, Google DeepMind now employs approximately 7,700 people as of February 2026.
What is Browser Agents?
AI agents and infrastructure for autonomously navigating web browsers—clicking, typing, scraping, and completing multi-step web tasks for testing and automation.
Browse all Browser Agents tools →