UI regression testing
Run realistic user journeys across browser, mobile, or desktop interfaces and capture where flows diverge, fail, or produce unexpected screens.
Independent tool overview
Gemini Computer Use at a glance
Gemini Computer Use is a preview Gemini API tool that lets supported models inspect screenshots and propose clicks, typing, scrolling, and other interface actions for browser, mobile, and desktop automation.
Visit the official Gemini Computer Use site ↗
- Current recommended model
- Gemini 3.7 Flash
- Original 2.5 model
- Legacy preview
- Environments
- Browser, mobile, and desktop on Gemini 3.x
- Execution
- Client-side; developers run the proposed actions
- Tool status
- Preview
- Billing
- Regular input and output token rates
- Reviewed
- August 30, 2026
Overview
What Gemini Computer Use is
Gemini Computer Use is a developer capability for building agents that operate graphical interfaces when a reliable API is unavailable or insufficient. An application sends a goal, current screenshot, environment type, and interaction history to Gemini; the model returns a proposed UI action; the application executes that action with a browser or device automation layer; and the updated screenshot is sent back for the next step. This loop can power form entry, web research, UI testing, legacy-system work, and cross-platform task automation.
This page originally covered the standalone Gemini 2.5 Computer Use preview model. That model is now labeled legacy. Google moved computer use into its general Gemini 3.x models, and the current documentation recommends the generally available `gemini-3.7-flash` model. Gemini 3.x expands the tool from browser-only use to browser, mobile, and desktop, adds human-readable action intents, configurable safety policies, and optional prompt-injection detection. The computer-use tool itself remains Preview, so important workflows need tight scopes, sandboxing, least-privilege accounts, confirmation before consequential actions, and deterministic verification.
Use cases
Who Gemini Computer Use is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Legacy application automation
Operate systems that expose important business workflows only through a graphical interface and do not have a usable API.
Repetitive data entry
Navigate predictable forms and transfer reviewed information between systems while retaining human confirmation for submissions and irreversible changes.
Cross-site research
Visit multiple sites, apply filters, collect visible details, and return structured findings when direct APIs or feeds do not cover the task.
Agent product prototypes
Evaluate whether screenshot-based control can handle a bounded workflow before investing in custom integrations or a broader automation platform.
Capabilities
Core Gemini Computer Use features
Screenshot-based perception
Uses screenshots and conversation history to interpret the current interface state instead of depending on a site-specific DOM or API schema.
Browser, mobile, and desktop environments
Gemini 3.x models expose action sets for three environment types, broadening the capability beyond the original browser-focused 2.5 model.
UI action generation
Can propose clicks, multi-clicks, pointer movement, scrolling, typing, key presses, drag-style interactions, and other predefined actions supported by the selected environment.
Normalized coordinates
Returns coordinates on a 0-to-999 scale so the client can translate actions to the actual screen or viewport dimensions.
Action intents
Gemini 3.x responses include an `intent` explaining why each action was selected, helping developers log and review the agent's immediate objective.
Safety decisions
An internal policy layer can classify proposed actions as allowed, requiring confirmation, or blocked before the client executes them.
Configurable safety policies
Gemini 3.x integrations can tune defined safety categories and overrides for the application rather than relying on one fixed action policy.
Prompt-injection detection
Offers opt-in screenshot scanning intended to identify adversarial instructions embedded in the interface the agent is viewing.
Continuous interaction loop
Preserves model and function responses across steps so the agent can plan and adapt through a multi-screen task.
Custom client execution
Works with the developer's chosen automation stack—such as Playwright for browsers—so the application controls isolation, credentials, action execution, screenshots, and logging.
Multiple model choices
Supports the recommended Gemini 3.7 Flash, cost-focused Gemini 3.5 Flash-Lite, previous Gemini 3.5 Flash, Gemini 3 Flash Preview, and the legacy Gemini 2.5 Computer Use model.
Gemini API ecosystem
Can be combined with Gemini reasoning, function calling, Search or Maps grounding, and application-defined tools when UI control is only one part of the workflow.
Process
How the Gemini Computer Use workflow works
Step 1
Choose a narrow task
Define allowed sites or applications, data classes, actions, completion criteria, time and step limits, and every point that requires a person to approve.
Step 2
Create an isolated environment
Use a sandboxed browser, device, or desktop session with a dedicated low-privilege account, restricted network access, controlled downloads, and no unrelated secrets.
Step 3
Start the interaction
Send the user goal, active screenshot, target environment, tool configuration, and safety settings to a supported Gemini model.
Step 4
Validate each proposed action
Check the action name, coordinates, intent, safety decision, current page, and local allowlists; pause for confirmation before submissions, messages, purchases, deletions, permissions, or other consequences.
Step 5
Execute and capture state
Run the accepted action through the client automation layer, wait for the interface to settle, take a fresh screenshot, and return the function result to Gemini.
Step 6
Verify completion
Stop at the configured boundary, confirm the final state with deterministic checks or source-system data, preserve an audit log, and clean up the isolated session and credentials.
Cost
Gemini Computer Use pricing and free plan
Computer Use has no separate tool fee on current Gemini 3.x models; it is billed as the model's regular input and output tokens, including screenshots, action history, and thinking tokens. The tool is not available on the free tier. Multi-step loops can consume far more tokens than a single prompt, and any connected Search or Maps grounding can add its own usage fees.
Gemini 3.7 Flash
$0.75 input · $3.75 output / 1M tokens
Current recommended GA model for Computer Use, at introductory standard pricing through December 31, 2026.
- Regular token billing; no separate computer-use surcharge
- Prices rise to $1.50 input and $7.50 output on January 1, 2027
- Output price includes thinking tokens
- Priority and other service modes have different rates
- Free model access does not make the Computer Use tool available on the free tier
Gemini 3.5 Flash-Lite
$0.30 input · $2.50 output / 1M tokens
Lower-cost GA option for high-volume, simpler computer-use workloads.
- Regular token billing
- Lower cost may trade off task reliability and long-horizon performance
- Benchmark on the exact workflow before scaling
Gemini 3.5 Flash
$1.50 input · $9 output / 1M tokens
Previous stable Gemini Flash model with built-in computer use.
- Regular token billing
- Gemini 3.7 Flash is the current recommended target
- Batch pricing is not a natural fit for interactive action loops
Gemini 2.5 Computer Use legacy
$1.25–$2.50 input · $10–$15 output / 1M tokens
Legacy browser-focused preview model retained for existing integrations.
- $1.25 input and $10 output at prompts up to 200K tokens
- $2.50 input and $15 output above 200K tokens
- No free tier
- New builds should prefer supported Gemini 3.x models
Pricing checked . Check current pricing at the source ↗
Assessment
Gemini Computer Use strengths and limitations
Where it stands out
- Works with graphical interfaces that lack a suitable API or stable structured integration.
- Gemini 3.x covers browser, mobile, and desktop rather than only browser control.
- The recommended Gemini 3.7 Flash model is GA even though the Computer Use tool remains Preview.
- Action intents make step-level logging and human review easier than opaque coordinate calls.
- Configurable safety decisions and confirmation states give applications a framework for consequential-action controls.
- Optional prompt-injection detection addresses a major risk specific to agents reading untrusted screens.
- Client-side execution lets teams choose their own sandbox, automation driver, credentials, network rules, and observability.
- A lower-cost Flash-Lite option supports experiments on high-volume but bounded workflows.
- The same Gemini stack can combine UI action with reasoning and structured tools.
What to consider
- Computer Use is a Preview capability and Google explicitly warns that it may contain errors and security vulnerabilities.
- The original Gemini 2.5 Computer Use model is legacy; new applications need a Gemini 3.x integration and may require response-schema and safety-flow changes.
- The API does not operate a computer for you: developers must build, secure, observe, and maintain the browser, mobile, or desktop execution environment and agent loop.
- Screenshot-based control is slower and more fragile than a stable API because layouts, pop-ups, animation, localization, ads, accessibility modes, and responsive breakpoints can move targets.
- Prompt injection can be hidden in websites, messages, documents, and images; opt-in detection reduces but does not eliminate that risk.
- Screenshots, typed values, history, and model responses can expose sensitive data unless the application redacts, isolates, and minimizes what the model and logs receive.
- The model can click the wrong target, misread state, repeat actions, enter incorrect data, or falsely conclude that a task succeeded.
- Important submissions, payments, messages, account changes, legal commitments, deletions, permissions, and access to sensitive data require explicit human control and final verification.
- Long action loops repeatedly send screenshots and history, making cost, latency, and context growth materially higher than one API call.
- Gemini 3.7 Flash's introductory price expires December 31, 2026, so production cost models must include the announced higher 2027 rates.
- Computer Use is not offered on the free tier even when the underlying Gemini model has free developer access.
- CAPTCHAs, bot controls, multifactor authentication, anti-automation policies, and website terms may block or prohibit automated interaction.
- Mobile and desktop automation introduces OS permissions, application isolation, update compatibility, and device-management concerns beyond browser automation.
- A direct API, database integration, deterministic script, or conventional test should be preferred whenever it can complete the task reliably and safely.
Compare
Gemini Computer Use alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Agents
ChatGPT Agent
ChatGPT Agent is a user-facing managed agent with its own computer and tools; it is easier for end users but less like a low-level developer API for embedding custom UI-control loops.
Explore ChatGPT Agent →Agents
Gemini Managed Agents
Gemini Managed Agents are a better fit when the priority is a hosted background agent and sandbox with remote tools rather than maintaining a screenshot-and-action runtime yourself.
Explore Gemini Managed Agents →Coding
Codex
Codex is better when the task can be completed through code, files, terminal tools, tests, or APIs; use visual computer control only for the portions that truly require a graphical interface.
Explore Codex →Questions
Gemini Computer Use FAQs
What is Gemini Computer Use?
It is a Gemini API tool for building agents that inspect screenshots and propose interface actions such as clicking, typing, scrolling, and key presses across supported browser, mobile, and desktop environments.
Is Gemini 2.5 Computer Use still current?
No. Google labels the standalone `gemini-2.5-computer-use-preview-10-2025` model as legacy. Current documentation recommends `gemini-3.7-flash`, where computer use is a built-in tool.
Does Google run the browser or device?
No. Your application receives action calls from the model and must execute them in a client-side browser, mobile, or desktop automation environment, capture the new screenshot, and continue the loop.
How much does Gemini Computer Use cost?
Current Gemini 3.x computer use is charged at the selected model's normal input and output token rates without a separate tool fee. The recommended Gemini 3.7 Flash is $0.75 input and $3.75 output per million tokens through December 31, 2026, then doubles to its announced standard rate.
Is there a free tier?
No. Google's pricing table says the Computer Use tool is not available on the free tier, even though some supported Gemini models offer free usage for other capabilities.
Can it control mobile and desktop apps?
Yes with supported Gemini 3.x models and a compatible client executor. The legacy Gemini 2.5 Computer Use model is browser-focused.
How does human confirmation work?
Gemini can return a safety decision requiring confirmation, but the application must implement the pause, show the user the exact proposed consequence, receive authorization, refresh state, and only then execute.
Is prompt-injection detection enough to make browser agents safe?
No. It is an additional signal, not a complete defense. Strong deployments also isolate sessions, restrict sites and actions, minimize credentials, verify destinations, block sensitive operations, and require human approval at consequential boundaries.
When should I avoid Computer Use?
Avoid it for critical decisions, unsupervised sensitive-data access, irreversible actions, or any workflow that a stable API or deterministic automation can perform more reliably and securely.
Bottom line
Our Gemini Computer Use verdict
Gemini Computer Use is promising infrastructure for the awkward last mile where software exposes work only through a screen. The move into Gemini 3.7 Flash brings broader environments, action intents, and better safety controls while lowering the current token rate. But it remains a preview building block, not an autonomous employee: the buyer still owns the sandbox, credentials, policy, confirmation UI, prompt-injection defenses, error recovery, and proof of completion. Start with read-only or reversible workflows, compare against a direct integration, and expand authority only after measured reliability and adversarial testing.
Visit Gemini Computer Use website ↗Similar Agents Tools
View all Agents AI tools →
Petri
Petri - Anthropic's open-source agentic system for evaluating LLM safety
Agent Skills
Agent Skills - Customize Claude with specialized capabilities
Agent Builder
Agent Builder - OpenAI's visual canvas for creating and versioning multi-agent workflows
FlowithOS
FlowithOS - An agentic workspace that connects your knowledge, creation, and execution in a single flow
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.