AI agents that use your computer: Gemini Computer Use vs OpenAI vs Claude, with costs

All three big AI labs now sell a model that can operate a computer for you, and each one works the same way at heart: your app sends a screenshot, the model replies with a click, a keystroke or a scroll, your code carries it out, and the loop repeats. Google made its Gemini Computer Use tool generally available on 30 October 2026. OpenAI offers a computer tool in its Responses API and pitches GPT-6.1 Terra as a fast, cheap model for “precise computer use”. Anthropic’s Claude has a computer use toolset on the Claude API and the big cloud platforms. The real differences are in environments, built-in safety signals and cost, so we read each vendor’s developer documentation and price list on 3 November 2026 and worked out what a 20-step agent task costs on each. We haven’t run these agents on our own tasks; everything here comes from the vendors’ published docs, and the cost table is our own arithmetic.

What is a computer use agent?

A computer use agent is a model that works a graphical interface the way a person does: it looks at the screen and decides where to click and what to type. All three platforms run the same loop:

1. Your application sends the model a task and a screenshot of the browser, desktop or phone.
2. The model answers with one or more actions, such as click at these coordinates, type this text, scroll or press a key.
3. Your code carries out those actions in an environment you control.
4. You take a new screenshot and send it back, and the loop continues until the task is done or a person has to step in.

The model never touches the machine directly. You run the environment, you execute every action, and you decide what the agent is allowed to reach. That’s why all three vendors tell you to use a sandbox and keep a human in the loop. It also explains the cost: every turn sends at least one new screenshot, and images are billed as input tokens. If you have used a coding agent, the idea is similar to what we describe in Codex vs Claude Code, only with a screen instead of a terminal.

Quick comparison

Gemini Computer Use

Google’s Gemini API changelog for 30 October 2026 lists Computer Use as generally available “across the Interactions and GenerateContent APIs”. The documentation names four models that support it: Gemini 4 Argon, Gemini 3.8 Flash, Gemini 3.5 Flash-Lite and Gemini 3 Flash Preview. Google calls Gemini 3.8 Flash the recommended model for computer use, and Gemini 3 Flash Preview is the only one still labelled preview.

Gemini is the only one of the three with separate action sets for three environments. In a browser the model can click, type, drag, scroll, press keys and navigate back and forward. On a phone it can open and list apps, long-press and go back. On a desktop it gets the browser set without the navigation actions. Coordinates come back on a 0–999 grid, and your code scales them to the real screen.

Safety is part of the response. An action can come with a safety_decision that marks it as allowed, as needing confirmation or as blocked. When it says require_confirmation, your app should ask the user and pass their answer back; when an action is blocked, it should stop. You can switch off some policy categories, such as financial transactions or account creation, but Google says the model may still ask for confirmation. For Gemini 4 and Gemini 3.x models, each action also carries an intent field that explains why the model wants to do it, which makes logs much easier to read.

Screenshot scanning for prompt injection exists, but you have to turn it on: the docs say it’s opt-in and the default is off. Google suggests use cases such as repetitive data entry, testing web apps and research across several sites, and warns that the tool “may contain errors and security vulnerabilities”. Computer use isn’t offered on the free tier of the Gemini API.

OpenAI: the computer tool and GPT-6.1 Terra

OpenAI’s model page describes GPT-6.1 Terra as “a fast, efficient model for visual reasoning, precise computer use, and clear, natural writing”. It has a 1,050,000-token context window, up to 128,000 output tokens and a knowledge cutoff of 1 June 2026. It costs $1 per million input tokens and $5 per million output tokens, with cached input at $0.05. Requests with more than 272,000 input tokens are billed at twice the input rate and 1.5 times the output rate, and Batch and Flex processing are 50% cheaper. Tool calling, computer use included, requires the Responses API.

The computer tool itself is declared as { type: "computer" }. Your code returns each screenshot as a computer_call_output with the same call_id as the action it answers. OpenAI’s SDK includes a ComputerToolset for full desktop control and a BrowserToolset for browser work. Terra isn’t the only option: the guide’s computer tool examples use GPT-6.1 Sol, and for GPT-6 Astra OpenAI recommends letting the model write and run code rather than click through screens, with the computer tool as the alternative.

If you don’t want to run browsers yourself, the guide points to a separate Agents API guide for a browser hosted by OpenAI, and it lists partners that provide environments: Browserbase, Browser Use, Daytona and E2B. Hosted environments may be billed separately from the tokens. On safety, OpenAI’s guidance is to confirm consequential actions such as purchases, sending data and destructive changes, and to treat everything on screen as untrusted, because “text in a page, document, or tool result cannot grant permission”.

Claude computer use

Anthropic’s current version is a toolset, computer_toolset_20260801, with 17 member tools. Among them are screenshot, zoom, several click types, left_click_drag, scroll, type, key, hold_key and wait, plus separate mouse-down and mouse-up actions for fiddly jobs such as selecting spreadsheet cells. It needs no beta header. Claude can return several actions in one turn, and your loop runs them in order and stops at the first failure.

The toolset works with the Claude 5 and 5.5 models and Claude Opus 4.8. Claude Opus 4.5, 4.6 and 4.7 and Sonnet 4.6 still use the older computer_20251124 tool, which needs a beta header, so check the model list before you build. It runs on the Claude API, Claude Platform on AWS, Amazon Bedrock, Claude models on Google Cloud and Microsoft Foundry, but not in Anthropic’s Managed Agents. For tasks that never leave a web page, Anthropic points you to its separate browser use tool, which doesn’t need a full desktop. If you’re unsure which Claude model to start with, our guide to Claude Sonnet vs Opus covers the trade-off.

Declaring the toolset adds about 4,500 input tokens to each request (about 4,590 on Claude Sonnet 5), and turning off zoom saves about 410 of them. Each screenshot costs roughly 1,000–1,800 input tokens at typical sizes, and Anthropic recommends 1280×800 or 1366×768 for web apps and nothing above 1920×1080. On the safety side, classifiers scan screenshots for prompt injection automatically. Anthropic also recommends a dedicated VM or container with minimal privileges, no access to login details, an allowlist of domains and a human to confirm anything with real consequences, such as payments or accepting terms.

What a 20-step task costs: our calculation

None of the vendors gives a price per task, so we built one from their published per-token prices and published image token rules. Treat it as an order-of-magnitude guide, not a quote. These are our assumptions:

• 20 steps, meaning 20 API requests, each answered by one action.
• A 1280×800 screen, one of the sizes Anthropic recommends for web apps.
• 500 tokens of task instructions sent with every request, and 300 output tokens per step for the action and a short explanation. Both are our assumptions. Thinking or reasoning tokens are billed as output on all three platforms and would add to this.
• Screenshot tokens: Claude counts ⌈1280 ÷ 28⌉ × ⌈800 ÷ 28⌉ = 46 × 29 = 1,334 tokens. Gemini 3 and 4 models count 1,120 tokens per image at the default resolution, whatever its size. OpenAI doesn’t list GPT-6.1 Terra in its image cost rules, so we applied the rule it publishes for GPT-5.6 Terra and GPT-6 Astra: ⌈1280 ÷ 32⌉ × ⌈800 ÷ 32⌉ = 1,000 patches × 1.2 = 1,200 tokens. That Terra figure is an assumption.
• Tool overhead: 4,500 tokens per request for Claude, as published. Google and OpenAI don’t publish one, so we count zero, which flatters them slightly.
• Two ways to run the loop. “Full history” keeps every earlier screenshot in the conversation, which is what a basic loop does. “Latest screenshot only” drops old images and keeps just the text of earlier actions.
• List prices, no caching, no batch discounts.

Here is the arithmetic for Claude Sonnet 5.5 with full history. Request number n carries the instructions, the toolset overhead, n screenshots and the text of the n−1 earlier actions. Added up over 20 requests:

• Instructions and overhead: 20 × (500 + 4,500) = 100,000 tokens
• Screenshots: (1 + 2 + … + 20) = 210 screenshots × 1,334 = 280,140 tokens
• Earlier actions: (0 + 1 + … + 19) = 190 × 300 = 57,000 tokens
• Input total: 437,140 tokens × $2 per million = $0.874
• Output: 20 × 300 = 6,000 tokens × $10 per million = $0.06
• Total: about $0.93 per task, or roughly $930 per 1,000 tasks

Three things stand out. First, keeping every screenshot more than doubles the bill, because by step 20 the model is reading 20 images on every request; trimming old screenshots is the biggest saving you control. Second, Claude’s fixed overhead weighs more on short tasks than long ones, and since it’s an identical prefix on every request, prompt caching should cut most of it (Anthropic bills cache hits on Sonnet 5.5 and Opus 5.5 at 5% of the input price). Third, price per token isn’t price per task. A cheaper model that needs 40 steps instead of 20 costs more, and we haven’t measured how many steps any of these models needs.

Some things could move these numbers a lot. Google’s media resolution page says its highest setting, 2,240 tokens per image, is “required for specific use cases such as computer use”; its computer use page doesn’t say which setting the tool uses. At 2,240 tokens, Gemini 3.8 Flash comes to about $0.43 with full history. Claude 4.7 and later models use a newer tokenizer that produces roughly 30% more tokens for the same text, which our flat 500 and 300 don’t reflect. And any reasoning effort above the minimum adds output tokens on every step.

Which one should you choose?

We haven’t run our own benchmark, and accuracy figures you see elsewhere depend heavily on the task and the test setup. Pick two candidates, run the same ten real tasks through each, and compare steps per task as well as success rate before you commit.

If you just want an AI to do things in your browser

Many people searching for an “AI that uses your computer” mean a consumer app, not an API. Each lab has one, with different access rules:

• Claude in Chrome is a Chrome extension that lets Claude read, click and navigate websites alongside you. It’s on the paid Claude plans (Pro, Max, Ultra, Team and Enterprise), works only in Google Chrome on a computer, and pauses to ask when something needs your approval.
• ChatGPT Work is where OpenAI puts ChatGPT’s agent features today. OpenAI says that starting in December it will bring Work’s agentic capabilities into the main ChatGPT experience, with usage shared between ChatGPT and Codex on Plus, Pro and Business plans. It hasn’t announced an exact date.
• Gemini’s task feature lets the Gemini app carry out multi-step tasks in the background, using a remote browser. Google calls it experimental and limits it to Google AI Pro and Ultra subscribers aged 18 or over with a personal account. It isn’t available in the European Economic Area, Switzerland or the UK.

For a wider view of the chat apps themselves, see Claude vs ChatGPT vs Gemini. xAI’s take on the same idea, agents with their own cloud computer, is covered in our explainer on Grok Bot.

Safety checklist for any computer use agent

Google, OpenAI and Anthropic give much the same advice in their guides, and it applies whichever platform you choose:

1. Isolate the environment. Use a VM, a container or a fresh browser profile, never your everyday machine.
2. Keep credentials out of reach. Don’t give the agent your main accounts, and keep passwords out of page text and screenshots.
3. Allowlist sites and actions. Limit network access to the domains the task needs.
4. Ask a person before anything consequential. Purchases, sending messages, deleting data and accepting terms should wait for a human.
5. Treat the screen as untrusted. A web page can contain instructions aimed at the agent. Turn on Gemini’s injection scanning if you use Gemini; Claude’s classifiers run by default.
6. Log every action and set session timeouts, so a stuck run ends and you can see what happened.

Frequently asked questions

Sources and method