Computer Use Tools Compared - Claude, OpenAI, and Gemini on Who Runs the Environment and Its Safety Controls

First Published:
Last Updated:

Computer use is a tool that shows a model the screen and lets it operate the mouse and keyboard. Anthropic's Claude, OpenAI, and Google's Gemini each offer these tools via their respective APIs. While all three companies use the term "computer use," who runs the screen and input environment, who applies confirmation before a risky action, and whether prompt injection detection is enabled by default differ by route.

This article sets out six routes offered by these three companies, not as step-by-step instructions, but as questions to ask when choosing. The questions are: 1) Who runs the execution environment? 2) What does the model see, and what does it return? 3) Where does confirmation live, and what is enabled by default? 4) Where does the data stay, and on which platforms is each tool offered? 5) Which versions stopped, and when?

In short, the OpenAI Agents API route is the only one of the six where the vendor runs the execution environment. In the remaining five routes, the user side runs the environment and executes the model's requests through their own code. The defaults for confirmation and detection do not point the same way across the three companies. Claude has prompt injection classification enabled by default, while Gemini's detection is disabled by default. OpenAI's Agents API requires approval for each new origin, but does not enforce confirmation before individual actions. This article lists these differences side by side without aligning them.

This article is based on materials reviewed on November 20, 2026 (Japan Standard Time). The status of these tools and the names of the models changed between July and November 2026 as well. The dates on which information was verified are included in the tables. This article does not cover pricing, performance comparisons, or which product is better.

Related articles on this site:

Table of Contents

  1. 1. Scope and Terms
  2. 2. Who Runs the Environment
  3. 3. What the Model Sees and Returns
  4. 4. Where the Confirmation Lives
  5. 5. Prompt Injection Defenses and Their Defaults
  6. 6. Data Retention and Where Each Tool Is Offered
  7. 7. Status, Versions, and What Has Been Shut Down
  8. 8. Comparison Table
  9. 9. Failure Modes
  10. 10. Frequently Asked Questions About Computer Use Tools
  11. 11. Summary
  12. 12. References

1. Scope and Terms

This chapter first sets out the readers of this article, its scope, the sources it read, and its terms. In short, its conclusions are the following ten points.

  • This article compares six routes across three companies. These are Claude's computer use and browser use, OpenAI's Responses API computer tool, code execution, and Agents API, and Gemini's Computer Use (Section 1.2).
  • Only in the OpenAI Agents API route does the vendor run the execution environment. In the other five routes, the user side runs the environment (Chapter 2).
  • Claude's computer use is not available in Claude Managed Agents. The same applies to its browser use (Section 2.1).
  • The information the models access primarily consists of screenshots and coordinates. In addition, Claude's browser use can read the page's accessibility tree. Gemini supports three environments: browser, mobile, and desktop (Chapter 3).
  • Gemini's safety_decision is the one that requests confirmation as a structure in the API response. With Claude and the OpenAI Responses API, the user's application applies confirmation (Chapter 4).
  • OpenAI's Agents API requests user approval for each new origin. However, it does not enforce confirmation before individual actions (Section 4.3).
  • Claude's injection classifier is enabled by default. Gemini's injection detection is disabled by default. No statement about detection settings was found in OpenAI's pages (Chapter 5).
  • Claude's computer use and browser use are eligible for zero data retention (ZDR), excluding Covered Models. OpenAI's Responses API is also subject to ZDR, with an exception for image input. The Agents API is not subject to ZDR (Chapter 6).
  • Claude's computer use came out of beta on the Claude API on August 19, 2026. Gemini's Computer Use reached General Availability (GA) on October 30, 2026 (Chapter 7).
  • Earlier versions and preview models are stopping. Claude's 5.5 series models return a 400 error for the earlier computer_20251124 on the Claude API and Google Cloud. OpenAI's computer-use-preview was shut down on July 23, 2026 (Chapter 7).

1.1 Intended Audience

This article is intended for the following audiences:

  • Developers creating applications or agents that delegate screen interactions to models, and who need to determine which vendor and which route to use.
  • Individuals responsible for checking the safety controls and data handling when using computer use in their organization.
  • Operations personnel who are currently using computer use and want to understand the impact of earlier versions and preview models stopping.

This article assumes that readers may hold the following views, and checks each of them against the sources:

  • With computer use, the vendor provides the screen environment.
  • The API takes care of confirmation before a risky action.
  • Prompt injection detection is enabled by default at every vendor.
  • The "computer use" offerings from the three companies refer to the same core functionality.

1.2 Scope and Exclusions

This article addresses six specific routes. This article uses these six as the fixed row names in its tables.

  • Claude computer use: The toolset computer_toolset_20260801.
  • Claude browser use: The toolset browser_toolset_20260801.
  • OpenAI Responses computer tool: The computer tool in the Responses API.
  • OpenAI code execution: A method in which the model writes code that is then executed in the user's environment.
  • OpenAI Agents API: The computer_use tool in the Agents API, and the openai_hosted environment.
  • Gemini Computer Use: The computer_use tool in the Gemini API.

This article covers the following aspects for each route: the operator of the environment, units of operation, environment types, confirmation mechanisms and their defaults, prompt injection defenses and their defaults, data retention, platforms, status and dates, earlier versions and shutdowns, and billing units.

This article does not cover:

  • Usage instructions and code examples. Declarations are kept to a minimum.
  • Procedures for migrating from earlier versions to the current version. This article covers only which earlier versions return a 400 error.
  • Comparisons of model performance, pricing, or which product is better.
  • Injection examples, methods for avoiding detection, or configurations that bypass approval. This article limits the scope to comparing whether security controls are present, whether they are enabled by default, and who applies them.
  • Comparisons of third-party hosts such as Browserbase and E2B. This article only lists their names.
  • What managed execution platforms hold regarding sessions and credentials. This site's Managed Agent Runtimes covers this.
  • Details regarding AWS's Amazon Bedrock AgentCore Browser. This site's What AgentCore Browser Keeps and Where Its Traffic Goes covers this.
  • The history of dates related to computer use. This site's Tool Use and Agent Protocol History and Timeline covers this as a timeline.

1.3 Date of Review and Materials Reviewed

This article was written after reviewing the following materials on November 20, 2026 (Japan Standard Time). In the body, this article uses the English names in parentheses.

Regarding Anthropic's Claude Platform, the following pages were reviewed:

  • Computer use tool (Claude computer use page)
  • Browser use tool (Claude browser use page)
  • Browser and computer use with the SDK toolsets (Claude SDK toolsets page)
  • Claude Platform release notes (Claude release notes)
  • API and data retention (Claude data retention page)

Regarding OpenAI's API documentation, the following pages were reviewed:

  • Computer use (OpenAI computer use guide)
  • Computer use integration recipes (OpenAI integration recipes)
  • Computer use for the Agents API (OpenAI Agents API computer use page)
  • Changelog (OpenAI changelog)
  • Deprecations (OpenAI deprecations page)
  • Data controls in the OpenAI platform (OpenAI data controls page)
  • Safety classifiers (OpenAI safety checks page)

Regarding Google's documentation, the following pages were reviewed:

  • Computer use for the Gemini API (Gemini computer use page)
  • Release notes for the Gemini API (Gemini release notes)
  • Gemini deprecations (Gemini deprecations page)
  • Zero data retention in the Gemini Developer API (Gemini ZDR page)
  • Computer use for the Gemini Enterprise Agent Platform (Agent Platform computer use page)

For AWS, this article read Anthropic Claude tool use in the Amazon Bedrock User Guide (Bedrock Claude tool use page) and the AgentCore Browser page in the Amazon Bedrock AgentCore Developer Guide (AgentCore Browser page).

For pages outside AWS, this article read the original content obtained using curl. The Gemini API pages redirected from ai.google.dev to aistudio.google.com. For the Gemini API pages, this article read the Markdown version rather than the HTML. For AWS, this article read the content through the search function of AWS Knowledge.

1.4 Terms, and Different Things with the Same Spelling

This article will consistently use the following terms:

  • Environment: The browser, virtual machine, container, or device whose screen and input the model operates.
  • Operator of the environment: The entity that runs the environment, updates the browsers and operating systems within it, and fulfills the model's requests. This article borrows the definition of the operator of the host from this site's WebAssembly Outside the Browser, dividing it into two categories: the vendor and the user side. When the user side rents a third-party host, this article calls it a third-party host.
  • Route: One of the six routes described in Section 1.2.
  • Confirmation: The act of a person approving an action before it is executed.
  • Detection: The mechanism for identifying injections within a screen or page.
  • Default: The value used when the user has not configured any specific settings.

The sources this article read use the same spelling for different things.

  • Computer use: In Claude, this refers to the toolset computer_toolset_20260801, as well as the earlier tools computer_20251124 and computer_20250124. In OpenAI, it refers to the computer tool type for the Responses API, the earlier tool type computer_use_preview, the model computer-use-preview that has been shut down, and the computer_use tool for the Agents API. In Gemini, it refers to the tool computer_use. Unless otherwise specified, this article uses "computer use" as a general term encompassing all six routes, identifying individual components by their identifiers.
  • Browser use: Claude's browser_toolset_20260801 is a toolset defined by Anthropic. OpenAI's SDK has a BrowserToolset class, through which the user's driver exposes operations as function tools. "Browser Use" is the name of a third-party library and a cloud service. In this article, "browser use" will be used exclusively to refer to Claude's browser_toolset_20260801, unless otherwise stated.
  • Toolset: Claude's computer_toolset_20260801 represents a type of tool for the API. Both Claude's SDK and OpenAI's SDK contain SDK classes whose names also include toolset. When referring to the SDK classes in this article, it will be explicitly stated that they are SDK classes.
  • Google Cloud: In Claude's documentation, "Google Cloud" refers to Claude Models on Google Cloud. As of November 20, 2026, the Claude computer use page lists it in two rows: Anthropic managed and Google managed. Where a source writes Google Cloud without distinguishing the two, this article also writes Google Cloud. In the Gemini documentation, Agent Platform refers to the Gemini Enterprise Agent Platform. These are two separate platforms.
  • GA: This article uses "GA" only when the documentation states generally available or GA. Following the term out of beta in the Claude release notes, this article describes Claude's change on August 19, 2026, as coming out of beta.

1.5 Terms for the State of the Sources

This article uses the following six terms to describe the state of the documentation. These terms, including their English spellings, are also used in the tables.

  • Stated: The documentation contains a statement regarding the matter.
  • Stated as not supported: The documentation explicitly states that the matter is not supported or cannot be used.
  • Documentation does not list: While a list or table exists, the specific item is not included. This does not necessarily indicate that the matter is unsupported or unavailable.
  • No statement found: Despite searching using different keywords and categories of documentation (as described in Section 1.3), no relevant statement could be found. This does not necessarily indicate that the matter is absent or impossible.
  • Sources disagree: The statements regarding the same matter in two different documents are contradictory. This article presents both statements, along with their sources, without designating either as incorrect.
  • Inference: This is an inference made by this article, and is not a statement found in any of the documentation. This article says in the same sentence that it is an inference.

All dates referenced in this article are presented with the year. Pages may be rewritten after the verification date.

2. Who Runs the Environment

This chapter shows who runs the execution environment for each of the six routes.

Who Runs the Computer Use Environment
Who Runs the Computer Use Environment

2.1 Claude: The User Side Runs It

The Claude computer use page describes the computer use toolset as follows:

The computer use tool is an Anthropic-defined client toolset: one {"type": "computer_toolset_20260801"} entry in tools gives Claude 17 member tools such as screenshot, left_click, type, and zoom, and your application runs every call in an environment you control. It isn’t currently available in Claude Managed Agents.

This means that the user side runs the execution environment. Claude simply returns model requests; it is the user's application that captures the screen and performs clicks. The Availability section on the same page also marks the Managed Agents row as Not available.

The same page describes the reference implementation as follows:

For security and isolation, the reference implementation runs all of this inside a Docker container with appropriate port mappings for viewing and interacting with the environment.

The Claude browser use page describes the same division of responsibility for browser use.

Your application runs every call against its own browser automation; nothing runs on Anthropic’s side.

Browser use is also not available in Claude Managed Agents. The Claude browser use page states The tool isn’t currently available in Claude Managed Agents.

Claude Managed Agents is a product in which Anthropic runs the agent loop and the container that executes tools. The fact that computer use is not available in that container overlaps with the scope covered by this site's Managed Agent Runtimes. This article treats computer use and browser use as routes where the user side runs the environment.

2.2 OpenAI: The Operator Differs Across Three Routes

The OpenAI computer use guide poses a question regarding the selection of a path, phrasing it as follows:

Choose a path based on what the model needs to see and who runs the environment

Regarding the two routes that use the Responses API directly, the same page states:

Use the Responses API directly when you need to manage the tool loop yourself. You provide the environment and execute the model’s requests.

These two routes are code execution and the computer tool. In code execution, the model writes code that uses libraries such as PyAutoGUI or Playwright, and that code is then executed in the user's environment. With the computer tool, the model returns structured mouse and keyboard commands, which the user's application then translates into input. In both, the user side runs the execution environment.

The same page describes GPT-6 Astra, stating:

For GPT-6 Astra, we recommend code execution. The computer tool remains supported as an alternative.

This article quotes this sentence only as the page's wording, and does not recommend either option.

The same page also includes a table outlining scenarios where the user does not run the environment themselves, instead relying on a third-party host. The table lists names such as Browserbase and Browser Use. Even in these cases, the code required to fulfill the model's requests resides on the user's side.

The third route is the Agents API. The OpenAI computer use guide points to the Agents API page for Run a browser hosted by OpenAI. The OpenAI changelog, dated September 29, 2026, states:

Added computer use to the Agents API. Agents can complete tasks in an OpenAI-hosted browser, with website access approvals and sign-in handled by your application.

The OpenAI Agents API computer use page instructs users to configure the browser by adding the computer_use tool to agent.tools and setting environment.type to openai_hosted. In this route, the vendor, OpenAI, runs the browser. However, the user's application handles approvals for website access and sign-in (Section 4.3).

2.3 Gemini: The User Side Runs It

The Gemini computer use page states the following regarding the execution environment:

Similar to function calling, you will need to implement the client-side execution environment to receive and execute the Computer Use actions.

In other words, the user side runs the execution environment. The Gemini computer use page says that the reference implementation includes a Docker-based sandbox. Similarly, the Agent Platform computer use page indicates that users will write code for client-side applications.

2.4 AWS: Bedrock Is a Platform, and AgentCore Browser Is a Vendor-Run Browser

AWS documentation addresses two distinct topics.

First, Amazon Bedrock is a platform that offers Claude's computer use. The Bedrock Claude tool use page states that Claude's two toolsets are generally available on the newer Claude models and require no beta value. Here too, the user side runs the execution environment.

Second, it introduces Amazon Bedrock AgentCore Browser. The AgentCore Browser page describes it as an isolated browser environment that allows agents to interact with web applications. Users can choose a browser that AWS manages (aws.browser.v1). This is an example where the vendor runs the execution environment, but it is not a tool in any of the three companies' model APIs. This article does not include AgentCore Browser in the comparison of the six routes; instead, it directs readers to this site's What AgentCore Browser Keeps and Where Its Traffic Goes.

2.5 Operators by Route

Table 1 lists the operators responsible for the execution environment for each of the six routes, along with the supporting text for that designation. The verification date is November 20, 2026.

RouteOperator of the EnvironmentSupporting Text
Claude computer useUser sideyour application runs every call in an environment you control (Claude computer use page)
Claude browser useUser sidenothing runs on Anthropic’s side (Claude browser use page)
OpenAI Responses computer toolUser side or a third-party hostYou provide the environment and execute the model’s requests. (OpenAI computer use guide)
OpenAI code executionUser side or a third-party hostSame as above
OpenAI Agents APIVendor (OpenAI)an OpenAI-hosted browser (OpenAI changelog, September 29, 2026)
Gemini Computer UseUser sideimplement the client-side execution environment (Gemini computer use page)

3. What the Model Sees and Returns

This chapter shows, for each route, what the model sees and in what unit it returns actions.

3.1 Screenshots and Coordinates

Claude's computer use involves manipulating the screen using screenshots and coordinates. The Claude computer use page lists 17 member tools, including screenshot, left_click, type, and zoom. Claude can return multiple actions in a single response. The Claude computer use page describes this as a "batch action," noting that the user's application executes these actions sequentially, rather than in parallel.

The OpenAI Responses computer tool also operates using screenshots and coordinates. The OpenAI integration recipes list the following actions that the computer tool can return: click, double_click, scroll, type, wait, keypress, drag, move, and screenshot. A single computer_call contains an ordered array of actions.

In OpenAI's code execution, the model does not return actions directly, but rather code that performs those actions. The OpenAI computer use guide states that a single call can combine actions, loops, and conditional branching.

Gemini's Computer Use also operates using screenshots and coordinates. The Gemini computer use page specifies that coordinates are represented as integers in the range of 0 to 999. Adjusting the coordinates to fit the screen size is the responsibility of the user's code. For Gemini 4 and Gemini 3.x models, each action carries an intent field that explains why the model chose that action.

3.2 Understanding Page Structure: Claude's Browser Use

Claude's browser use involves understanding the structure of a page, in addition to using screenshots. The Claude browser use page states:

Claude works with the page both through its structure (the accessibility tree, elements, forms, and tabs) and through screenshots and viewport coordinates.

The browser use toolset has 27 member tools by default, and enabling the optional members adds four more. The Claude computer use page notes that browser use is the closer fit for tasks that stay inside webpages. The Claude browser use page specifies that computer use should be used for tasks that require access to the entire desktop.

3.3 Environment Types

Gemini's Computer Use functionality operates in three environments. The Gemini computer use page lists ENVIRONMENT_BROWSER, ENVIRONMENT_MOBILE, and ENVIRONMENT_DESKTOP, specifying the mobile environment as being for Android.

The Claude computer use page describes the environments used for the reference implementation, listing a virtual display (Xvfb), a desktop environment on Linux, and applications such as Firefox. This describes the configuration for the reference implementation, and is not a list of supported operating systems.

The OpenAI Agents API route uses a browser that OpenAI runs. The example on the OpenAI Agents API computer use page sets environment.desktop.enabled to true.

3.4 Unit of Operation Per Route

Table 2 lists, for each route, what the model observes and what it returns. The verification date is November 20, 2026.

RouteModel ObservationsModel ReturnsMultiple Operations Per Response
Claude computer useScreenshotsCalls to the 17 member toolsStated (batch action)
Claude browser useAccessibility tree, elements, forms, tabs, screenshotsCalls to member tools: 27 by default, plus 4 optional onesStated
OpenAI Responses computer toolScreenshotsArray of actions within computer_callStated
OpenAI code executionScreenshots and the results of code executionCode to perform operationsStated (single script)
OpenAI Agents APIScreen of the browser run by OpenAIOperations within an Agents API sessionNo statement found
Gemini Computer UseScreenshotsCalls to functions with coordinates (0-999) and intentNo statement found

4. Where the Confirmation Lives

This chapter clarifies whether the confirmation process for potentially risky operations resides in the API response, in the user's application, or in the vendor's environment. It also shows the default for each.

Where the Confirmation Lives and What Is On by Default
Where the Confirmation Lives and What Is On by Default

4.1 Claude: Your Executor Applies It

The Claude computer use page, under the Security considerations section, lists the following as one measure to reduce risk:

Asking a human to confirm decisions that might result in meaningful real-world consequences and any tasks requiring affirmative consent, such as accepting cookies, completing financial transactions, or agreeing to terms of service.

This is a precaution, not a mechanism by which the API stops an action. It is the user's application that applies confirmation.

The Claude browser use page is more specific about where to put confirmation.

Have a human confirm consequential actions and anything that requires affirmative consent (purchasing, modifying accounts, messaging, and accepting terms), and make that check in your executor before each call, because one turn can carry several.

Because a single response may contain multiple operations, confirmation goes before each call, rather than once per response. The Claude release notes, dated October 7, 2026, state that the Python SDK version 1.12.0 and the TypeScript SDK version 0.132.0 include beta classes for both tools. These classes allow the SDK to execute approval callbacks provided by the user. Here too, the decision on confirmation rests with the user's code.

4.2 OpenAI Responses API: Your Application Applies It

OpenAI integration recipes describe confirmation processes as follows:

Apply confirmation and consent rules in your application and execution environment. Decide whether to execute a request, pause for approval, or hand control to the user. The model’s request to act is not user permission.

The same section states that in responses involving multiple operations, the process should stop before the initial operation requiring confirmation. In code execution, where a single script may perform numerous operations, it is necessary to impose permissions on both the auxiliary functions being exposed and the execution environment.

The OpenAI computer use guide lists four measures for safe operation: Restrict the environment., Treat screen content as untrusted., Confirm consequential actions., and Bound and verify the run. All of these measures are imposed in the user's application and the execution environment.

A computer_call in the Responses API has a pending_safety_checks field. In the examples provided in the OpenAI integration recipes, this field is an empty array. The OpenAI computer use guide describes the example code as follows:

Pending API safety checks stop this sample; add a human approval callback only when your application can handle them.

The pages this article read have no statement on when this field contains a value (No statement found). This article states only that the field exists, and does not describe it as a mechanism by which the API requests confirmation.

4.3 OpenAI Agents API: Approval for Each Origin

The OpenAI Agents API computer use page describes approvals in the browser that the vendor runs as follows:

The browser requires the user’s approval before accessing each new website origin, including public websites. Enabling network access does not approve these requests.

In this route, approval for each new origin is required by default. Enabling network access does not count as approval. The user's application receives each approval request and answers approve, deny, or cancel.

However, the same page states the scope of this approval as follows:

Origin approval does not enforce confirmation before individual actions

The next sentence is guidance for applications that must guarantee confirmation before purchases or destructive changes.

If your application must guarantee confirmation before purchases, destructive changes, or other consequential actions, restrict the hosted browser to resources that cannot perform them, or use a browser runtime you control.

This article summarizes this route as follows: approval for each origin is required by default, but confirmation before individual actions is not enforced. The steps for origin approval and the handling of sessions are covered in this site's Managed Agent Runtimes.

The OpenAI changelog states that on October 30, 2026, automatic review (auto_review) was added to the Agents API. When the session's review field is set to auto_review, a model approves or denies the supported tool and permission requests. The OpenAI Agents API computer use page further explains:

Enabling automatic review also preserves browser origin approvals and sign-in requests. Continue handling them in your application.

In other words, even with automatic review enabled, origin approvals and sign-in requests remain with the user's application.

4.4 Gemini: API Response May Request Confirmation

The Gemini computer use page states that model responses may include a safety_decision value. This value indicates whether an action is permitted as standard, requires user confirmation (require_confirmation), or should be blocked. The same page describes the confirmation process as follows:

If the safety_decision is require_confirmation, prompt the end user. If the user confirms, set safety_acknowledgement in function_result.

In Gemini, the API response structurally returns whether confirmation is required. The user's application should read this value and prompt the user, then send a confirmation back via safety_acknowledgement.

The Gemini computer use page lists seven built-in categories that form the basis of the safety policies used to determine when confirmation is required: FINANCIAL_TRANSACTIONS, SENSITIVE_DATA_MODIFICATION, COMMUNICATION_TOOL, ACCOUNT_CREATION, DATA_MODIFICATION, USER_CONSENT_MANAGEMENT, and LEGAL_TERMS_AND_AGREEMENTS. Users can disable these categories using the disabled_safety_policies setting. Because a setting to disable them exists, this article reads the categories as active when the user configures nothing. This is an inference. However, the same page also states:

Safety overrides indicate your preference, but the model may still return a safety_decision with require_confirmation in some cases. Your application should always implement safety decision handling regardless of configured overrides.

Even if categories are disabled, the API may still return require_confirmation. Therefore, regardless of the user's settings, the user's application must handle this value.

The Gemini computer use page also provides examples of excluding predefined functions with excluded_predefined_functions and adding a custom function that hands control back to the user.

4.5 Confirmation and Defaults by Route

Table 3 lists where confirmation lives and its default for each route. The verification date is November 20, 2026.

RouteWhere Confirmation LivesConfirmation the API Applies by DefaultState of the Sources
Claude computer useYour executorNone. Confirmation is described as a precaution.Stated (Security considerations section)
Claude browser useYour executor, before each callNone. Confirmation is described as a precaution.Stated
OpenAI Responses computer toolThe user's application and execution environmentNone. A pending_safety_checks field is present.Stated. Conditions for a value to be entered in the field: No statement found.
OpenAI code executionThe helpers and runtime exposed to the codeNoneStated
OpenAI Agents APIThe vendor's browser requests approval for each origin, and the user's application answersApproval for each new origin is required by default. Confirmation before individual actions is not enforced.Stated
Gemini Computer Usesafety_decision in the API response, and the user's applicationIncludes seven built-in categories. Can be disabled using disabled_safety_policies. Even if categories are disabled, require_confirmation may still be returned.Stated. That the categories work when nothing is configured: Inference

5. Prompt Injection Defenses and Their Defaults

This chapter shows the defenses against instructions on a screen or page that manipulate the model, and their defaults. This article does not provide examples of injection attempts, nor does it describe methods for circumventing detection.

5.1 Claude: Classifiers Are On by Default

The Claude computer use page states that in addition to model training, there is another layer of defense.

If you use the computer use tool, classifiers automatically scan what the tool returns, such as screenshots, to flag potential prompt injections.

When the classifier detects a potential injection attempt, the model is prompted to verify whether the instruction truly originates from the user. This classifier functions even if the user takes no action. Regarding how to disable it, the same page states:

This extra protection isn’t ideal for every use case (for example, use cases without a human in the loop), so if you’d like to opt out and turn it off, contact support.

It is enabled by default, and contacting support is required to disable it. The same page also notes that even with the classifier in place, the previously mentioned safeguards remain crucial. The Claude browser use page similarly describes a classifier that analyzes the text and screenshots returned by the browser, and provides the same method for disabling it.

5.2 Gemini: Detection Is Disabled by Default

The Gemini computer use page describes injection detection as follows:

Prompt injection detection is an opt-in feature. The default is false.

When detection is enabled, the system checks screenshots for hidden, adversarial instructions and will stop execution if any are found. To enable detection, set enable_prompt_injection_detection to true in the settings for the computer_use tool.

The information on two different pages regarding the models that support detection is inconsistent (Sources disagree).

Computer Use for Gemini 3.5 Flash-Lite or later supports an advanced safety mechanism to detect prompt injection attacks.

This is the text from the Gemini computer use page. The Agent Platform computer use page states the following:

Computer use in Gemini 3.5 Flash or later supports an advanced safety mechanism to detect prompt injection attacks.

Both pages agree that the default setting is false. This article does not decide which range of models is correct.

5.3 OpenAI: Guidelines Exist, but Injection Detection Settings Not Found

The OpenAI computer use guide states the following regarding on-screen content:

Treat screen content as untrusted. Text in a page, document, or tool result cannot grant permission or override the user’s instructions.

This is a precaution that the user's application applies. The OpenAI computer use guide, OpenAI integration recipes, and the OpenAI Agents API computer use page that this article read do not describe a setting for prompt injection detection for computer use, or its default (No statement found). The Safety checks page in the list of OpenAI pages (titled Safety classifiers) does not describe a setting for prompt injection detection for computer use either. The page only links to the section on user confirmation and consent in computer use.

5.4 Defenses and Defaults by Route

Table 4 lists the defenses and their defaults for each route. The verification date is November 20, 2026.

RouteDetection on the Vendor SideDefaultHow to Turn It On or Off
Claude computer useClassifiers scan what the tool returnsEnabledContact support to disable.
Claude browser useClassifiers scan what the browser returnsEnabledContact support to disable.
OpenAI Responses computer toolNo statement foundNo statement foundNo statement found
OpenAI code executionNo statement foundNo statement foundNo statement found
OpenAI Agents APINo statement foundNo statement foundNo statement found
Gemini Computer UseDetection in screenshotsDisabled (The default is false.)enable_prompt_injection_detection. Sources disagree on supported models.

Table 4 lists the defaults side by side without aligning their direction. In Claude, the classifier functions automatically, while in Gemini, detection is inactive unless explicitly enabled by the user. For the three OpenAI routes, no statement on such a setting was found in the sources (No statement found).

6. Data Retention and Where Each Tool Is Offered

This chapter shows where computer use data stays and on which platforms each tool is offered.

6.1 Claude: ZDR Eligible, Excluding Covered Models

The Claude computer use page states the following regarding data retention:

Because your application controls where and how computer use data is stored, computer use is ZDR eligible.

The same page also states that screenshots, mouse interactions, keyboard inputs, and files remain in the user's environment, while Anthropic processes screenshot and interaction requests within API calls. The "Data handling" table in the "Availability" section lists ZDR as Eligible (excludes Covered Models). The Claude data retention page identifies Claude Fable 5.2, Claude Mythos 5.2, and similar models as Covered Models, which require 30-day retention, so ZDR is not available for them unless Anthropic expressly authorizes it.

The Claude browser use page also lists browser use as ZDR eligible, excluding Covered Models. The handling of HIPAA differs between the two tools. The Claude computer use page lists HIPAA as Eligible, while the Claude browser use page lists it as Not eligible.

6.2 OpenAI: The Responses API Is ZDR Eligible, and the Agents API Is Not

The table on the OpenAI data controls page indicates, for each endpoint, whether it is subject to ZDR. /v1/responses is marked as Yes, see below for limitations, while /v1/agents is marked as No.

The same page specifically mentions computer use in relation to image input.

Images and files may be uploaded as inputs to /v1/responses (including when using the Computer Use tool), /v1/chat/completions, and /v1/images.

The next sentence states that a CSAM classifier scans image and file inputs, and that even if ZDR is enabled, potentially concerning images will be retained for human review.

The OpenAI computer use guide states the following regarding data:

The examples set store to false, but that does not remove every form of data retention.

The same section also states that when using third-party hosts, those hosts have their own data policies.

6.3 Gemini: No Statement Names Computer Use

The Gemini ZDR page does not contain any statements that specifically refer to computer use. Instead, it states, as a general provision for the Gemini Developer API:

  • Even for paid services, Google logs prompts and responses for a limited time to detect prohibited use. If zero data retention must be guaranteed, use Vertex AI.
  • The Interactions API stores conversation state by default. To disable this, set the store value to false.

An example on the Gemini computer use page uses the Interactions API. Taking the two together, this article reads that, to avoid keeping state in Gemini's computer use, the store value needs checking. This is an inference.

6.4 Platforms

The supportedPlatforms field on the Claude computer use page lists Claude API, Amazon Bedrock, Amazon Bedrock (bedrock-runtime), Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed), Claude Models on Google Cloud (Google managed), and Microsoft Foundry, all marked as stable. The Claude browser use page also lists the same seven platforms. The Claude release notes entry dated October 22, 2026, announced the Anthropic managed platform. Which models are available on which platforms varies by model and platform. This article does not reproduce those combinations, instead directing readers to the "Availability" sections of each respective page.

The Claude release notes state that the two toolsets became available on Google Cloud on August 20, 2026. The Bedrock Claude tool use page states that the two toolsets are generally available on the newer Claude models and require no beta value.

OpenAI's three routes are routes of the OpenAI API. In the pages it read, this article has not checked statements about other platforms.

Gemini's Computer Use is available on two platforms: the Gemini API and the Gemini Enterprise Agent Platform. The Gemini release notes name the Interactions API and the GenerateContent API as the scope of the GA. The list of models supported on the Agent Platform computer use page is more extensive than the list provided on the Gemini computer use page.

6.5 Data by Route

Table 5 lists how data is handled for each route. The verification date is November 20, 2026.

RouteWhere Screen and Action Data StaysZDRState of the Sources
Claude computer useUser's environment. Anthropic processes this within API calls.Eligible. Excludes Covered Models.Stated
Claude browser useUser's environment. The content returned is part of the API request.Eligible. Excludes Covered Models.Stated
OpenAI Responses computer toolUser's environment or a third-party host./v1/responses is included. An exception exists for image inputs.Stated
OpenAI code executionUser's environment or a third-party host.Same as above.Stated
OpenAI Agents APIEnvironment run by OpenAI./v1/agents is excluded.Stated
Gemini Computer UseUser's environment.No specific mention of "computer use" is found.No statement found (Gemini ZDR page).

In this article's scope, the billing units are as follows (specific amounts are not listed). Claude's two toolsets are billed according to the standard tool use charges, and declaring a toolset adds input tokens. Screenshots are counted as image inputs. The Agent Platform computer use page states that computer use is billed at the same rates and with the same SKUs as Gemini. The OpenAI computer use guide states that screenshots are counted as image inputs, and that a hosted environment may be billed separately.

7. Status, Versions, and What Has Been Shut Down

This chapter shows the status and dates for each route, and the versions that stopped. The history of these dates is documented in this site's Tool Use and Agent Protocol History and Timeline, which presents the information as a timeline. This article will only list the dates relevant to the selection process.

7.1 Claude: The Version That Came Out of Beta, and Earlier Versions That Return 400

The Claude release notes state that on October 22, 2024, computer use tools for the new Claude Sonnet 3.5 were added to the API. An entry dated August 19, 2026, reads as follows:

The computer use tool is out of beta on the Claude API as the computer_toolset_20260801 toolset: no beta header, batch actions (several actions in one turn), zoom enabled by default, and per-member configuration through configs. Earlier beta versions remain available.

Following the term out of beta in the Claude release notes, this article describes this change as coming out of beta. On the same day, the browser use toolset, browser_toolset_20260801, was released.

The earlier versions, computer_20251124 and computer_20250124, remain in beta. The Claude computer use page states the following about them:

Two earlier versions of the computer use tool are available in beta.

However, the new models do not accept these earlier versions. The Claude release notes indicate that on September 22, 2026, for Claude Opus 5.5 and on September 28, 2026, for Claude Sonnet 5.5, the Claude API and Google Cloud do not accept computer_20251124. For Claude Haiku 5.5, released on October 7, 2026, computer_20250124 returns a 400 error. An entry dated November 9, 2026, discusses Claude Fable 5.2 and Claude Mythos 5.2, stating:

A request that declares the earlier computer_20251124 tool returns a 400 error on every platform except Amazon Bedrock, where the earlier tool still works.

The Claude computer use page indicates that on Amazon Bedrock, Claude Fable 5.2, Claude Mythos 5.2, Claude Opus 5.5, Claude Sonnet 5.5, and Claude Haiku 5.5 all accept computer_20251124 with the beta header. This article does not detail the steps for migrating from the earlier versions.

7.2 OpenAI: The Preview That Was Shut Down, and the GA computer Tool

The OpenAI changelog states that on March 11, 2025, OpenAI released the Responses API and built-in tools, including computer use, and announced the computer-use-preview model. The entry dated March 5, 2026, states that GPT-5.4 supports built-in computer use through the Responses API's computer tool.

The Past deprecations table on the OpenAI deprecations page gives July 23, 2026, as the shutdown date of a deprecation announced on April 22, 2026.

The table lists computer-use-preview-2025-03-11 and computer-use-preview, with gpt-5.6-terra as the replacement model. The same page notes that models with "preview" in their names may be retired with a short notice period of approximately two weeks, citing computer-use-preview as an example.

The migration table in OpenAI integration recipes lists both preview and GA integrations. The model transitions from computer-use-preview to gpt-6.1-sol, the tool type changes from computer_use_preview to computer, and one action per call is replaced by an actions array.

Regarding the Agents API, the OpenAI changelog states that it was released as a public beta on September 10, 2026, and that computer use was added on September 29, 2026. The entry dated October 30, 2026, also refers to the Agents API as a public beta.

7.3 Gemini: From Preview to GA

The Gemini release notes state that the Gemini 2.5 Computer Use Preview was released on October 7, 2025. They also state that a public preview of the Computer Use tool in Gemini 3.5 Flash was released on June 24, 2026. The section dated October 30, 2026, reads as follows:

Released the Computer Use tool to general availability (GA) across the Interactions and GenerateContent APIs.

The table on the Gemini deprecations page lists July 28, 2026, as the shutdown date of gemini-2.5-computer-use-preview-10-2025, and gemini-3.8-flash as the replacement model.

7.4 Status and Dates by Route

Table 6 lists the status and dates for each route. The verification date is November 20, 2026.

RouteStatus (Verification Date)Key DatesStopped/Not Accepted
Claude computer usestable (7 platforms)Came out of beta on the Claude API on August 19, 2026. Available on Google Cloud on August 20, 2026.The 5.5 models and Fable and Mythos 5.2 return a 400 error for computer_20251124 on the Claude API and Google Cloud. Amazon Bedrock accepts it.
Claude browser usestableReleased on August 19, 2026.No statement found.
OpenAI Responses computer toolGA integrationSupported by GPT-5.4 on March 5, 2026.computer-use-preview was shut down on July 23, 2026.
OpenAI code executionNo statement foundThe OpenAI computer use guide recommends it for GPT-6 Astra. GPT-6 Astra was released on September 3, 2026.No statement found.
OpenAI Agents APIPublic betaAdded computer use on September 29, 2026. Added auto_review on October 30, 2026.No statement found.
Gemini Computer UseGAPublic preview on June 24, 2026. GA on October 30, 2026.gemini-2.5-computer-use-preview-10-2025 was shut down on July 28, 2026.

8. Comparison Table

This chapter summarizes the content from Chapters 2 through 7 in a single table.

8.1 Comparison of Six Routes

Table 7 lists six routes across nine columns. The verification date is November 20, 2026. In the table, None means that the API applies no such mechanism by default.

RouteOperator of the EnvironmentUnit of OperationEnvironmentConfirmation and DefaultDetection and DefaultDataBilling UnitStatus and DatePlatforms
Claude computer useUser sideScreenshots and coordinates, 17 member toolsUser's desktopYour executor. None from the API by default.Classifier. Enabled by default.ZDR eligible. Excludes Covered Models.Tokens for tool use and toolset definition, image input.stable. Came out of beta on August 19, 2026.Seven. Not available in Managed Agents.
Claude browser useUser sideAccessibility tree and screenshots, 27 member tools by defaultUser's browserYour executor, before each call. None from the API by default.Classifier. Enabled by default.ZDR eligible. Excludes Covered Models. Not eligible for HIPAA.Same as above.stable. Released on August 19, 2026.Seven. Not available in Managed Agents.
OpenAI Responses computer toolUser side or third-party hostScreenshots and coordinates, array of actionsUser's browser or desktopUser's application. None from the API by default.No statement found./v1/responses is ZDR eligible. Exception for image input.Image input.GA integration. computer-use-preview was shut down on July 23, 2026.OpenAI's API.
OpenAI code executionUser side or third-party hostCode that performs the actionsUser's browser or desktopExposed helpers and runtime. None from the API by default.No statement found.Same as above.Same as above.No statement found.OpenAI's API.
OpenAI Agents APIVendor (OpenAI)Operations in the browser OpenAI runsBrowser run by OpenAIApproval for each new origin is required by default. Confirmation before individual actions is not enforced.No statement found./v1/agents is not ZDR eligible.A hosted environment may be billed separately.Public beta. Computer use added on September 29, 2026.OpenAI's API.
Gemini Computer UseUser sideCoordinates from 0 to 999 and intentBrowser, mobile, desktopsafety_decision. Includes 7 built-in categories.Detection is opt-in. Default is false.No specific mention of computer use.Same pricing and SKU as Gemini.GA on October 30, 2026.Gemini API, Agent Platform.

8.2 Where the Sources Disagree, and Where No Statement Was Found

The following are the matters this article marks as Sources disagree:

  • Models targeted for Gemini injection detection. The Gemini computer use page states Gemini 3.5 Flash-Lite and later, while the Agent Platform computer use page states Gemini 3.5 Flash and later (Section 5.2).

The following are the matters this article marks as No statement found:

  • The conditions under which OpenAI's pending_safety_checks field receives a value (Section 4.2).
  • The injection detection settings and default values for OpenAI's three routes, specifically for computer use (Section 5.3).
  • Specific mentions of computer use on the Gemini ZDR page (Section 6.3).
  • Whether OpenAI's Agents API and Gemini can return multiple operations in a single response (Section 3.4).

In addition, OpenAI's documentation uses two different names for the model that replaces computer-use-preview. The OpenAI deprecations page lists gpt-5.6-terra, while the OpenAI integration recipes migration table lists gpt-6.1-sol. This article does not call this Sources disagree. The former is referenced in a notice dated April 22, 2026, while the latter is the current migration table.

9. Failure Modes

This chapter lists common misunderstandings that can occur when selecting and using the six routes.

  • Confusing origin approval in the OpenAI Agents API with confirmation before individual actions. The OpenAI Agents API computer use page states that origin approval does not enforce confirmation before individual actions (Section 4.3).
  • Assuming that confirmation is unnecessary because Claude's classifier is enabled by default. The Claude computer use page clarifies that even with the classifier enabled, the previously mentioned safeguards remain important (Section 5.1).
  • Believing that Gemini's detection is enabled by default. The Gemini computer use page states that the default is false (Section 5.2).
  • Assuming that removing Gemini's safety policy categories will prevent confirmation requests. The Gemini computer use page explains that even if categories are removed, require_confirmation may still be returned (Section 4.4).
  • Sending computer_20251124 to the Claude API with Claude's 5.5 series models. The Claude release notes state that this will return a 400 error (Section 7.1).
  • Directly moving earlier integrations running on Amazon Bedrock to the Claude API. While Amazon Bedrock accepts the earlier version with the 5.5 and 5.2 models, the Claude API and Google Cloud do not. For the 5.2 models, the Claude release notes state that a 400 error is returned on every platform except Amazon Bedrock (Section 7.1).
  • Maintaining integrations that rely on computer-use-preview. The OpenAI deprecations page states that it was shut down on July 23, 2026 (Section 7.2).
  • Assuming that setting store to false will ensure that no data is retained. The OpenAI computer use guide clarifies that this does not eliminate all data retention (Section 6.2).
  • In Claude's browser use, delegating URL inspection to the API. The Claude browser use page states:

the API doesn’t filter the URLs Claude opens, so it can’t reject one for you.

  • Believing that Claude's computer use can be used in Claude Managed Agents. The Claude computer use page states that this is not possible (Section 2.1).

10. Frequently Asked Questions About Computer Use Tools

Q1. Does the vendor supply the environment for computer use?

Of the six routes, only the OpenAI Agents API route uses an environment run by the vendor. In this route, OpenAI provides the browser. For Claude's two toolsets, the OpenAI Responses computer tool and code execution, and Gemini's Computer Use, the user side runs the environment and executes the model's requests. In the two OpenAI routes, the user side can also rent a third-party host.

Q2. Can Claude's computer use be used with Claude Managed Agents?

No, it cannot. The Claude computer use page states It isn’t currently available in Claude Managed Agents., and the Availability section marks Managed Agents as Not available. The Claude browser use page similarly states the same regarding browser use.

Q3. Does the API enforce confirmation before a risky action?

The requirement varies depending on the route. With Gemini's Computer Use, the API response's safety_decision can return require_confirmation, prompting the user's application to ask the user. With OpenAI's Agents API, approval for each new origin is required by default, but confirmation before individual actions is not enforced. In Claude's two toolsets, and in the OpenAI Responses computer tool and code execution, it is the user's application that applies confirmation.

Q4. Is prompt injection detection enabled by default?

It varies by company. Claude's classifier is enabled by default, and disabling it requires contacting support. Gemini's detection is disabled by default and can be enabled using enable_prompt_injection_detection. For OpenAI's three routes, the pages this article read have no statement on detection settings.

Q5. Is Claude's computer_20251124 still usable?

It depends on the model and the platform. The Claude computer use page lists computer_20251124 as an earlier beta version, and lists Claude Fable 5.1, Claude Opus 5, Claude Opus 4.5, and others as models that can use it. On the Claude API and Google Cloud, Claude Opus 5.5, Claude Sonnet 5.5, Claude Haiku 5.5, Claude Fable 5.2, and Claude Mythos 5.2 return a 400 error for computer_20251124. However, on Amazon Bedrock, these models still accept computer_20251124 with the beta header. For Claude Fable 5.2 and Claude Mythos 5.2, the Claude release notes entry dated November 9, 2026, states that a 400 error is returned on every platform except Amazon Bedrock.

Q6. Is OpenAI's computer-use-preview still available?

No, it is not. The OpenAI deprecations page states that access to models, including computer-use-preview, was shut down on July 23, 2026. Current integrations use the computer tool in the Responses API.

Q7. Does computer use data fall under ZDR?

It varies depending on the route. Claude's computer use and browser use are subject to ZDR, excluding Covered Models. OpenAI's /v1/responses is subject to ZDR, with an exception for image inputs that undergo CSAM screening. OpenAI's /v1/agents is not subject to ZDR. The Gemini ZDR page does not specifically mention computer use, and states that the Interactions API stores state by default.

Q8. What is the difference between Claude's browser use and computer use?

Browser use refers to a set of tools for working within a browser, including the ability to read accessibility trees in addition to taking screenshots. Computer use, on the other hand, involves operating the entire desktop using only screenshots and coordinates. The Claude computer use page states that for tasks that stay inside webpages, browser use is the closer fit. HIPAA also differs: computer use is Eligible, while browser use is Not eligible.

11. Summary

The tools for computer use are selected based on the following five questions:

  • Operator of the environment: Only in the OpenAI Agents API route does the vendor run the environment. In the other five routes, the user side runs the environment and executes the model's requests.
  • What the model sees: Primarily, the model receives screenshots and coordinates. Claude's browser use also reads the accessibility tree, while Gemini supports three environments: browser, mobile, and desktop.
  • Confirmation: Gemini requests confirmation in the API response, while OpenAI's Agents API requires approval for each new origin. With Claude and the OpenAI Responses API, the user's application applies confirmation.
  • Default detection: Claude's classifier is enabled by default. Gemini's detection is disabled by default. No statement on the setting was found in OpenAI's pages. The defaults do not point the same way across the three companies.
  • Versions: Claude's 5.5 series models return a 400 error for the earlier version on the Claude API and Google Cloud. OpenAI's computer-use-preview was shut down on July 23, 2026, and Gemini's 2.5 preview on July 28, 2026.

12. References



References:
Tech Blog with curated related content

Written by Hidekazu Konishi