Repository navigation
gemini computer use vs claude computer use
This series has compared a dead product to a living tool before. This page is the other kind: both sides exist, both are documented, and both were verified against the vendors' own pages for this comparison on 2026-09-03. The occasion for it is Google's side of the fence settling down. Project Mariner, the consumer experiment, shut down on 4 May 2026, and what remains at Google for builders is exactly what Anthropic has offered all along: a computer use capability in the API, where your code owns the environment and the model supplies the next action. Two vendors, one architecture, real differences in the details.
Disclosure: this page sits on the wiki of invisible_playwright_mcp, an open-source agent in the same category. It appears in exactly one labeled aside near the end and nowhere else. Everything about the two vendors traces to their live documentation, quoted or paraphrased with dates.
Both tools run the same loop. Your application captures a screenshot and sends it with the task; the model returns an action; your code executes that action against a display you control; a fresh screenshot goes back; repeat until done. Neither vendor hosts the browser or the desktop for you. That is the deal in both cases: the capability is a building block, and the environment, the guardrails and the runtime are yours to build.
The dialects differ in shape. Gemini returns a suggested action as a function call with coordinates normalized to a 0-999 grid that you scale to your actual viewport. Claude's current toolset exposes 17 member tools - screenshot, the click family, drag, scroll, type, key, hold_key, wait, and a zoom action that captures a region at full resolution - with pixel coordinates against the screenshot you sent.
Google. The capability began as a dedicated model:
gemini-2.5-computer-use-preview-10-2025, introduced 7 October 2025 and
built on Gemini 2.5 Pro. It has since become a tool that mainline models
call. The docs today recommend gemini-3.8-flash and list several other
Gemini 3-family models, with the 2.5 preview model kept as the legacy
entry. Two things worth reading closely: the tool rides current flagship
models rather than a frozen specialist, and the docs still carry the
caveat that "as a Preview capability, Computer Use may contain errors
and security vulnerabilities" - their words, worth pricing in.
Anthropic. The current toolset version is
computer_toolset_20260801, which needs no beta header on the Claude API
and Google Cloud and is supported by the current model families,
including Claude Fable 5.1, Mythos 5.1, Opus 5 and Sonnet 5; some cloud
platforms carry it as beta. Older models (the Opus and Sonnet 4.x lines)
use the earlier computer_20251124 tool version, which does require a
beta header. The toolset is generally available rather than preview.
If you want a one-line status difference: Anthropic's tool has crossed into GA on its home platform; Google's is newer to the tool form and still labels the capability Preview.
Google's docs describe three targets: browser control as the primary and most mature case, mobile with Android-optimized behavior including app launching, and desktop via OS-level cursor commands. The original 2.5-era blog was explicit that the model was "primarily optimized for web browsers" and "not yet optimized for desktop OS-level control"; the desktop support has grown since, but the browser-first center of gravity remains visible in the docs.
Anthropic's tool is environment-agnostic by construction: it emits generic display actions, and whatever you can render to a screenshot - a browser in a container, a full desktop VM - is fair game. Anthropic publishes a reference container with a virtual display and browser as a starting point. In practice most deployments of either tool are browser deployments, which is why this page lives in a browser-agent wiki.
Google's launch material claims leading results for browser control on Online-Mind2Web, WebVoyager and AndroidWorld, and cites 70%+ accuracy at around 225 seconds latency on Browserbase's Online-Mind2Web harness. Anthropic's computer use documentation does not publish a comparable benchmark table on the tool page, so this wiki has no like-for-like number to set beside Google's, and we are not going to manufacture one. Treat each vendor's figures as that vendor's claim on that vendor's chosen harness, and if the decision matters, run both against your own task; an afternoon of real runs beats every table in this category.
Both are standard token-priced API calls, and in both the dominant cost is the same: screenshots. Anthropic's docs are usefully concrete here - roughly 1,000 to 1,800 input tokens per screenshot, a stricter per-image limit once a request carries more than 20 images, and for the current toolset a resolution ceiling of 2576 pixels on the long edge. Google's computer use docs do not publish per-screenshot token figures in the same way, so budget by measuring a real session. Either way, a long agentic browse is mostly paying to look at the screen again, and the loop's step count matters more than the per-token rate.
The two postures rhyme. Gemini's tool takes per-category safety configuration - categories such as financial transactions, communication tools and account creation can be set to require explicit user confirmation - and offers prompt injection detection as an opt-in flag. Google's launch material also describes a per-step safety service assessing each proposed action, and lists categories of action the model is trained to refuse outright. Claude's toolset ships prompt injection classifiers that flag suspect content in screenshots and steer the model to ask for confirmation, and the docs recommend dedicated VMs or containers with minimal privileges, domain allowlists, keeping credentials away from the model, and human confirmation for consequential actions.
The practical reading is the same for both: these tools are built to act under supervision inside an environment you have deliberately bounded, and if your plan involves handing one your main browser profile and walking away, neither vendor's design agrees with your plan.
| Axis | Gemini computer use | Claude computer use |
|---|---|---|
| Status on 2026-09-03 | Tool available; capability labeled Preview in docs | Current toolset GA, no beta header on Claude API |
| Current naming | Computer use tool; gemini-3.8-flash recommended, gemini-2.5-computer-use-preview-10-2025 legacy |
computer_toolset_20260801; computer_20251124 for older models |
| Action encoding | Function call, 0-999 normalized coordinates | 17 member tools, pixel coordinates, includes zoom |
| Environment focus | Browser first; Android-optimized mobile; desktop cursor commands | Whatever you render; reference container published |
| Published benchmark claims | Online-Mind2Web, WebVoyager, AndroidWorld leadership | None on the tool page |
| Screenshot cost guidance | Not published per screenshot | ~1,000-1,800 tokens each, documented limits |
| Prompt injection handling | Opt-in detection flag | Classifiers on by design, confirmation steering |
| Consumer sibling | Gemini Agent; auto browse in Chrome | Claude in Chrome extension |
An aside from the maintainers, labeled as such: our project, invisible_playwright_mcp, is not a third column of this table - it is an already-assembled agent rather than an API building block, it drives its own source-patched Firefox, and it can sit under either vendor's models via OpenRouter or an MCP assistant. If you were reading this page because you want the loop without building the loop, that is the shelf it sits on, next to the others surveyed in Open-source computer-use agents. We make it, so audit that claim like everything else here.
Pick Gemini's tool if your agents already live on Google's stack, your tasks are browser-and-Android shaped, and you want the newest mainline models carrying the capability. Pick Claude's if you want the GA contract, the finer-grained action vocabulary, and documentation that prices its own screenshots. Both choices leave every hard operational question - the environment, the egress, the pacing, what the site sees - on your desk. Those questions have their own pages here: what a page can tell about its visitor and when Claude computer use reads as a bot, which applies nearly unchanged to Gemini-driven environments.
Which is better, Gemini or Claude computer use? Neither on paper. Google publishes stronger browser benchmark claims; Anthropic ships the GA contract and more operational detail. Run both on your task.
Is Gemini computer use generally available? The tool is available via the Gemini API, AI Studio and Vertex AI, and the docs still label Computer Use a Preview capability. Anthropic's current toolset is GA on the Claude API.
What model do I use for Gemini computer use? The docs recommend
gemini-3.8-flash today; gemini-2.5-computer-use-preview-10-2025 is
the legacy dedicated model.
Do these replace Project Mariner or Operator? They are the developer halves that outlived both consumer products; the consumer replacements are covered on Project Mariner is gone and Is OpenAI Operator still available?.
Can either drive my real browser? They emit actions; what gets driven is whatever environment you provide. The vendors' consumer siblings (auto browse in Chrome, Claude in Chrome) are the drive-your-own-browser products.
Do they solve captchas or guarantee access to sites? No. Both vendors' safety material points the opposite direction, and sites remain free to challenge any visitor.
See also: OpenAI Operator vs Claude computer use for the earlier chapter of this series, Project Mariner is gone: what replaced it for Google's consumer arc, and Open-source computer-use agents for the pre-assembled open versions of this loop.
- Gemini API: Computer use documentation, fetched 2026-09-03: model naming and recommendations, the action loop, normalized coordinates, environment support, safety categories, the opt-in prompt injection flag, and the Preview caveat quoted above.
- Google: Introducing the Gemini 2.5 Computer Use model, fetched 2026-09-03: the 7 October 2025 introduction, benchmark claims, the Browserbase harness figure, the browser-first optimization quotes, and the per-step safety service.
- Anthropic: Computer use tool documentation, fetched 2026-09-03: toolset versions, model support and beta-header rules, the 17 member tools, screenshot token figures and image limits, prompt injection classifiers, and the security recommendations.
- TechSpot on the Project Mariner shutdown, fetched 2026-09-03, for the 4 May 2026 date.
- Anthropic: Claude in Chrome general availability, surfaced via search 2026-09-03, for the consumer-sibling row.
Refereed by the maintainers of invisible_playwright_mcp, an open-source agent that competes with both tools compared here. Our one paragraph is labeled; the rest is the vendors' own documentation, cited so you can check our whistle-blowing against the rulebook.
- OpenAI Operator alternatives
- Open-source Operator-style agents
- Is OpenAI Operator still available?
- OpenAI Operator vs Claude computer use
- browser-use alternatives
- Choosing an AI browser agent
- Open-source AI browser agents
- Open-source computer-use agents
- What is an AI web agent?
- AI browser agents vs traditional scraping
- Cloud browser infrastructure for AI agents, explained
- Browserbase alternatives
- Firecrawl vs an AI browser agent
- Skyvern alternatives
- Stagehand vs browser-use
- Project Mariner is gone: what replaced it
- Manus alternatives
- Gemini computer use vs Claude computer use
- A stealth browser MCP, reviewed honestly by its own wiki
- AI browser vs AI browser agent: which one do you want?
- AI browser agent vs RPA: which one fits the job
- AI browser agent vs n8n, Zapier and Make
- Vercel agent-browser alternatives, compared honestly
- What is an agentic browser? Definition and the two kinds
- Open-source agentic browsers: the three layers, compared
- Choosing an MCP server for browser automation: four axes
- Stealth MCP servers compared: Camoufox, nodriver, Patchright
- Playwright MCP alternatives, and the three you don't need
- Autonomous browser agents: the four rungs of autonomy
- What is actually free in the AI browser agent stack
- browser-use on GitHub: what the repo actually gives you
- Playwright MCP vs Chrome DevTools MCP: different jobs
- How to choose among MCP servers: a map by category
- Which MCP servers are worth adding to Claude Code
- MCP on GitHub: finding servers and judging them fast
- MCP vs an API: the decision, and what the wrapper costs
- MCP alternatives: when the protocol is the wrong shape
- Why does my AI agent get blocked?
- The timing signal AI agents give off
- Agent retry loops trip rate limits, not fingerprints
- Claude computer use detected as a bot
- browser-use getting blocked: what you can and cannot change
- Playwright MCP session blocked: four causes, four fixes
- Playwright MCP and captchas: what actually gets you past
- Cloudflare and a browser MCP server: what is being read
- Can an AI agent solve a captcha? The honest answer
- Getting an AI agent to fill out forms
- The best model for an MCP browser agent, and what it really costs
- Browser problem or model problem?
- How to use a Playwright MCP server with Claude Code
- Extracting data to a CSV with an AI agent
- Monitoring a page for changes with an AI agent
- How to let Claude Desktop control a browser
- How to add a browser to Cursor as an MCP server
- Using an AI agent to hunt for apartments
- Getting website data into Google Sheets with an AI agent
- Using an AI agent to download invoices from portals
- AI agents for web research
- Using an AI agent to test your own website
- How to add a browser to Cline as an MCP server
- Posting to social media with an AI agent
- Posting to Facebook with an AI agent
- Posting to Instagram with an AI agent
- Posting to X with an AI agent
- Automating LinkedIn posts: read this first
- Appointment bots: what they are and what an agent can legitimately do
- Track prices across sites with an AI agent
- Build a lead list with an AI browser agent
- Run an AI browser agent on a schedule
- AI browser agent with a local LLM: what changes
- Should you log your AI agent into your accounts?
- How to write a task an AI browser agent can follow
- Move data between two web apps with an AI agent
- The MCP server
- How the tools are shaped, and why
- Playwright MCP vs the Playwright CLI: which fits when
- Playwright MCP: browser is already in use, and the fix
- Playwright MCP best practices: four decisions that matter
- Playwright MCP with a proxy, and the three leaks it leaves
- A browser MCP server in GitHub Copilot: setup and limits
- Using a browser MCP server for web scraping: the pattern
- Which LLM for browser automation: the four properties
- How to build a browser agent, and what to take instead
- Getting an AI agent to log into a website: three routes
- MCP tools, resources and prompts: who controls each
- How many MCP tools is too many? The context arithmetic
- How to build an MCP server: the decisions, not the scaffold
- Local or remote MCP server: what changes, and what does not
- Writing an MCP client in Python: the thirty-line version
- Self-hosted AI agent: what one actually costs to run
- How long an AI browser agent takes per step, measured
- Text, HTML, snapshot or screenshot: what the agent should read
- Giving an AI browser agent a stopping condition
- Keeping an AI browser agent out of destructive actions
- Why did the AI agent click the wrong thing
- When the page changes under the AI agent
- Running one AI agent task across a list of sites
- Seeing a page as it appears in another country
- Getting data out of a dashboard with no export button
- Two browsers in one session: main and support
- Finding the dead links on a site with an AI agent
- Filling a CRM record from a company's website
- One form submission per spreadsheet row, with an AI agent
- Dated screenshots of a page as evidence
- Checking order and delivery status with an AI agent
- Reading a PDF that opens inside the browser
- Summarising a long page or thread with an AI agent
- Collecting every image on a page with its caption
- Collecting event and course listings with an AI agent
- Cancelling a subscription with an AI agent
- What an AI agent can and cannot do inside an iframe
- Shadow DOM and an AI agent: you can click it, you cannot read it
- How to add a browser to Codex as an MCP server
- What a page snapshot costs, per control
- How to let Gemini CLI use a browser
- Native selects and the ones that only look like selects
- Clicking by selector or by coordinates
- How long the agent waits before it gives up
- What a second browser costs
- Uploading a file with an AI agent, and why this one cannot
- Watching the agent work, and when it is worth it
- When not to use an AI browser agent
- Agent or script: deciding once instead of every time
- Using the keyboard instead of the mouse
- Secrets in an agent task: where they end up
- What an agent run should log
- Deduplicating what an AI agent collects
- Normalising values across sites
- Validating an AI agent's output
- Reading a table with an AI agent
- Driving a site's own search and filters
- The task works headed and fails headless
- How to let the AdaL CLI use a browser