Computer use is a model tool that takes a screenshot, decides on a click, type, or scroll action, and hands that action back to your code to execute. This article compares three specific implementations: Anthropic's computer_use tool, which works across several current Claude models rather than one fixed release (this article uses Claude Opus 4.8 for pricing and code examples, with Claude Sonnet 5 noted as the cheaper tier), Gemini 3.5 Flash's computer_use tool, and GPT-5.5's computer tool.
This is a comparison of the developer-facing model APIs, not consumer agent products. Claude Cowork, ChatGPT Agent, and Google's consumer browser-agent features are separate products built on top of (or alongside) these APIs, and several competing articles blur that line. We keep the two layers separate throughout.
Quick comparison
If you want the short answer before the detail, here it is.
What "computer use" actually means (and what it isn't)
All three vendors implement the same basic loop. You send a request with the computer-use tool enabled, the model looks at the current screenshot and returns an action, your code executes that action in a real environment, and you send back a new screenshot as the next turn's input. Anthropic calls the repeated cycle "the agent loop." Google's docs describe the same four steps, sending a request, receiving a suggested action, executing it, then capturing a new screenshot and sending it back. OpenAI's built-in loop follows the identical pattern, sending a task, inspecting the returned computer_call, running the actions, capturing the updated screen, and repeating until the model stops asking for more actions.
What differs across vendors is the layer this API tool sits under. Each company also ships a consumer-facing agent product that is not the same thing as the raw API.
Is Claude computer use the same as Claude Cowork?
No. Claude Cowork is Anthropic's consumer product, a research preview launched for Pro and Max subscribers, currently macOS-only for screen-based control (Cowork itself is available on Windows), single-device, single-session, with no enterprise or on-premise deployment. The computer_use tool described in Anthropic's developer docs is a separate, model-agnostic API tool that you call directly through the Messages API. Developers building their own agents use the API tool, not Cowork.
What happened to OpenAI Operator?
OpenAI Operator was merged into ChatGPT Agent on July 17, 2025, and the standalone Operator site was deprecated. ChatGPT Agent is the consumer product, combining Operator's browsing capabilities with deep research and code execution. The computer tool documented in OpenAI's API guide is the underlying Responses API tool developers call directly, and it's a separate layer from ChatGPT Agent.
Google's consumer-facing browser agent features are a similar case. Gemini 3.5 Flash's computer_use tool, called through client.interactions.create(), is the raw API layer developers integrate into their own harness. This article covers the three API tools, not any vendor's packaged consumer agent.
Does computer use cost extra per screenshot or action?
No. None of the three vendors charge a separate metered fee for computer use. All three bill it at their standard input/output token rates for the underlying model, with the computer-use tool definition adding a small, fixed amount of token overhead on top of whatever the screenshots and responses themselves cost.
Anthropic's pricing docs state directly that computer use "follows the standard tool use pricing," and using the tool adds 466 to 499 tokens to the system prompt plus 735 input tokens for the tool definition itself on Claude 4.x models. Google's pricing page lists computer use as "charged as regular tokens per model pricing," with no separate line item. OpenAI's model listing includes computer use as one of GPT-5.5's supported tools alongside functions, web search, and file search, with no separate metered price disclosed anywhere in its official docs.
Computer use pricing comparison
Claude's documentation is the only one of the three that gives a concrete per-screenshot token estimate, roughly 1,000 to 1,800 input tokens per screenshot in a long agent loop, which is why Anthropic recommends prompt-caching breakpoints and pruning old screenshots in batches. This figure is Claude-specific. Neither Google's nor OpenAI's official docs publish an equivalent per-screenshot estimate, so it should not be extrapolated to Gemini 3.5 Flash or GPT-5.5.
Feature and action comparison
Which one has the most detailed action set?
Gemini 3.5 Flash. It has the largest documented action list and the only environment-typed action sets, where browser, mobile, and desktop each expose a different set of actions (mobile adds open_app, list_apps, and long_press, for example). It's also the only one of the three with a documented intent field explaining the model's reasoning for each action.
Which one can zoom into small UI text?
Only Claude, and only with the computer_20251124 tool version. Its zoom action lets the model view a specific screen region at full resolution, useful for reading file names in a sidebar, tab titles, status-bar text, or button labels that aren't legible at the screenshot's default resolution. Neither Gemini 3.5 Flash's nor GPT-5.5's official action tables list an equivalent zoom or region-inspection action.
Code comparison: the same task in three APIs
No vendor publishes a direct side-by-side "same task, same target site" comparison across all three products. The three snippets below are drawn, trimmed, straight from each vendor's own official quick-start documentation. This is an illustrative shape comparison, not a benchmark claim.
Claude computer use example
# Python, Anthropic Messages API (beta), computer_20251124 tool
# Source: Anthropic computer use docs, Quick start (Python tab)
client = anthropic.Anthropic()
response = client.beta.messages.create(
model="claude-opus-4-8", # or another compatible model
max_tokens=1024,
tools=[
{
"type": "computer_20251124",
"name": "computer",
"display_width_px": 1024,
"display_height_px": 768,
"display_number": 1,
},
{"type": "text_editor_20250728", "name": "str_replace_based_edit_tool"},
{"type": "bash_20250124", "name": "bash"},
],
messages=[{"role": "user", "content": "Save a picture of a cat to my desktop."}],
betas=["computer-use-2025-11-24"],
)
print(response)
Computer use is in beta and requires the anthropic-beta: computer-use-2025-11-24 header (or the equivalent betas=[...] parameter shown above) on Claude Sonnet 5, Claude Opus 4.8, Claude Opus 4.7, Claude Opus 4.6, Claude Sonnet 4.6, and Claude Opus 4.5. The example uses Claude Opus 4.8 because that's the model shown in Anthropic's own snippet, but the tool itself is model-agnostic across that list, not tied to one release.
Gemini 3.5 Flash computer use example
# Python, google-genai client, client.interactions.create(), computer_use tool
# Source: Gemini API computer use docs, minimal example
from google import genai
client = genai.Client()
interaction = client.interactions.create(
model="gemini-3.5-flash",
input="Search for 'Gemini API' on Google.",
tools=[{"type": "computer_use", "environment": "browser"}]
)
print(interaction)
Gemini's docs also show a fuller example that enables prompt-injection scanning:
interaction = client.interactions.create(
model='gemini-3.5-flash',
input="Find a flight from SF to Hawaii on Jun 30th, coming back on Jul 6th",
tools=[
{
"type": "computer_use",
"environment": "browser",
"enable_prompt_injection_detection": True
}
]
)
The environment field ("browser", "mobile", or "desktop") is required and changes the available action set. Returned coordinates are normalized to a 0-999 range, so your application has to denormalize them against the real viewport before executing the action:
def denormalize_x(x: int, screen_width: int) -> int:
"""Convert normalized x coordinate (0-1000) to actual pixel coordinate."""
return int(x / 1000 * screen_width)
GPT-5.5 computer use example
# Python, OpenAI Responses API, built-in computer tool
# Source: OpenAI computer use guide, "Send the first request"
from openai import OpenAI
client = OpenAI()
response = client.responses.create(
model="gpt-5.5",
tools=[{"type": "computer"}],
input="Check whether the Filters panel is open. If it is not open, click Show filters. Then type penguin in the search box. Use the computer tool for UI interaction.",
)
print(response.output)
The first turn often returns a computer_call asking for a screenshot before the model commits to any UI action, which OpenAI's docs note is normal. You send the screenshot back using previous_response_id and a computer_call_output:
def send_computer_screenshot(response, call_id, screenshot_base64):
return client.responses.create(
model="gpt-5.5",
tools=[{"type": "computer"}],
previous_response_id=response.id,
input=[
{
"type": "computer_call_output",
"call_id": call_id,
"output": {
"type": "computer_screenshot",
"image_url": f"data:image/png;base64,{screenshot_base64}",
"detail": "original",
},
}
],
)
GPT-5.5 documents three harness options rather than one fixed shape, the built-in computer tool shown above, a custom harness (Playwright, Selenium, VNC, or MCP) that you drive through normal tool calling, or a code-execution harness where the model writes and runs short scripts and can move between visual and DOM-based interaction. Claude and Gemini each expose one canonical tool type, while GPT-5.5 treats computer use as one of three interchangeable integration paths.
Benchmarks: how accurate is each model?
No vendor publishes a first-party "error rate" for its own computer-use tool. Every accuracy number here comes from a third-party or independent benchmark, mainly OSWorld-Verified. There is also no single independent benchmark that has run all three exact named products (a specific Claude model, Gemini 3.5 Flash, and GPT-5.5) head-to-head. That's a genuine gap in the available data, not something we can paper over with mismatched numbers.
OSWorld-Verified scores compared
OpenAI's own announcement states GPT-5.5 "reaches 78.7%" on OSWorld-Verified, and its own comparison table separately cites Claude Opus 4.7 (not 4.8) at 78.0% on the same benchmark. BenchLM's independent leaderboard, as of July 20, 2026, ranks Claude Opus 4.8 first at 85.2, GPT-5.4 second at 79.2, and Claude Opus 4.6 third at 76.9.
BenchLM's leaderboard covers exactly five models, Claude Opus 4.8, GPT-5.4, Claude Opus 4.6, Qwen3.7 Plus, and Claude Opus 4.5. Neither GPT-5.5 nor any Gemini model appears on it. So while Claude Opus 4.8's 85.2 is the strongest independently-verified figure in this comparison, it is not a same-benchmark-run result against GPT-5.5 or Gemini 3.5 Flash specifically, and shouldn't be read as one.
Is Gemini 3.5 Flash's benchmark score verified?
No. The 78.4% figure comes from a third party citing DeepMind's model page, and that DeepMind page returned a poor-quality fetch when we tried to verify it directly, so it could not be independently confirmed. Treat it as reported, not as an Anthropic- or OpenAI-style vendor-confirmed figure. None of Google's own official computer-use or Gemini 3.5 Flash documentation that we found states a directly quotable OSWorld-Verified score for the model.
Safety and prompt injection handling
Claude runs automatic classifiers on screenshots to flag potential prompt injections. When a classifier flags something, it steers the model toward asking for user confirmation before the next action, and Anthropic notes this protection can be turned off by contacting support for use cases without a human in the loop. Anthropic's own docs also recommend using a dedicated sandboxed VM, avoiding access to sensitive login data, and limiting internet access to an allowlist of domains.
Gemini 3.5 Flash returns a safety_decision field on individual actions, classifying each one as allowed, require_confirmation, or blocked. Built-in safety-policy categories cover things like financial transactions, sensitive data modification, autonomous account creation, and accepting terms of service, and you can override select policies per request. Prompt injection detection, which scans screenshot pixels for hidden adversarial instructions, is opt-in.
GPT-5.5's guidance is procedural rather than a structured field on each action. OpenAI's docs recommend running computer use in an isolated browser or VM, keeping a human in the loop for high-impact actions, treating page content as untrusted input, and deciding upfront which sites and actions the agent is allowed to reach. No machine-readable per-action safety classification equivalent to Gemini's safety_decision is documented for the built-in computer tool.
Which one should you use?
Choose Claude's computer use tool if: you need a dedicated zoom action for fine-grained UI inspection, you're already building on Claude/Anthropic's tools, or you want the model with the strongest independently-verified OSWorld-Verified track record (with the caveat that this is Opus 4.8/4.6 on BenchLM's leaderboard, not a same-run comparison against GPT-5.5 or Gemini 3.5 Flash).
Choose Gemini 3.5 Flash's computer use tool if: you need environment-typed action sets with separate browser, mobile, and desktop configurations, you want per-step reasoning transparency via the
intentfield, or cost-per-token is the deciding factor ($1.50 input / $9.00 output vs. $5+ input for the other two).
Choose GPT-5.5's computer tool if: you want flexibility across three harness options (built-in tool, custom Playwright/Selenium/VNC/MCP harness, or code-execution harness) rather than one fixed API shape, or you want the one OSWorld-Verified score in this comparison that's fully vendor-confirmed via a primary source (78.7%, from OpenAI's own announcement).
FAQ
Do Claude, Gemini, and GPT charge extra for computer use?
No. All three bill computer use at their standard per-token input/output rates for the underlying model. Claude's docs disclose a small fixed token overhead for the system prompt (466-499 tokens) and tool definition (735 input tokens on Claude 4.x models). Gemini and GPT-5.5 disclose no equivalent overhead figure. None of the three charges a separate per-screenshot or per-action fee.
Is Claude computer use the same as Claude Cowork or ChatGPT Agent the same as OpenAI Operator?
No to both. Claude Cowork and ChatGPT Agent are consumer-facing products, and the computer_use and computer tools are the raw APIs developers call directly. OpenAI Operator specifically was merged into ChatGPT Agent on July 17, 2025, and the standalone Operator site was deprecated, but the underlying computer tool in the Responses API is a separate integration point from either product.
Which model scores highest on OSWorld-Verified?
It depends on the verification tier you're comfortable with. Claude Opus 4.8 leads on independently-verified numbers, scoring 85.2 on BenchLM's leaderboard. But BenchLM's leaderboard doesn't include GPT-5.5 or any Gemini model, so this isn't a same-benchmark-run comparison against either of the other two products in this article. GPT-5.5's 78.7% is the one figure here confirmed directly by its own vendor.
Can I use these computer-use tools without building my own sandbox?
No. All three are schema-defined tools that return actions, you still have to provide the execution environment. Claude needs a VM or container (Anthropic publishes a Docker-based reference implementation), Gemini 3.5 Flash needs a browser, mobile, or desktop harness (Google publishes its own Docker-based sandbox reference implementation), and GPT-5.5 needs a Playwright/Selenium/VNC/MCP harness or a code-execution harness that you build yourself, since no dedicated OpenAI reference-implementation repository was found alongside Anthropic's and Google's.