Gemini 2.5 Computer Use Is Now Legacy—and Gemini 3.6 Takes Over UI Control

Google has not turned Gemini 2.5 Computer Use into a generally available standalone product. As of August 13, 2026, it remains listed as a legacy preview, while Google’s current Computer Use documentation recommends Gemini 3.6 Flash for new agents that control browser, mobile and desktop interfaces.

That distinction changes the practical meaning of the 2025 release. Gemini 2.5 introduced developers to Google’s screenshot-driven interface agent, but it is now a compatibility option rather than the platform’s preferred foundation; the capability itself has moved into newer general-purpose Flash models.

What Google originally released

Gemini 2.5 Computer Use debuted on October 7, 2025, as a specialized model built on the visual understanding and reasoning capabilities of Gemini 2.5 Pro. It entered public preview through the Gemini API in Google AI Studio and Vertex AI, with browser interaction as its primary target rather than complete desktop operating-system control.

The agent did not receive unrestricted access to a computer. An application supplied a user request, a screenshot and recent action history; the model then proposed an interface operation such as clicking or typing. Client-side software executed that operation and returned an updated screenshot and URL, creating a repeated perception-and-action loop.

This architecture mattered because many workflows cannot be completed through a clean API. Forms, menus, filters, authenticated pages and older business applications often expose only a graphical interface. The original Google DeepMind announcement described browser-focused operation, support for human confirmation on certain actions and safeguards intended to stop dangerous or prohibited behavior.

Why the 2.5 model is no longer the default choice

The important update is architectural, not merely a new version number. Computer use began as a separate Gemini 2.5 preview model, but Google subsequently integrated the tool into its broader Flash line. The current model list labels Gemini 2.5 Computer Use as “Legacy Preview,” identifies Gemini 3.5 Flash as a previous stable option and places Gemini 3.6 Flash in the recommended position.

Gemini 3.6 Flash supports browser, mobile and desktop environments rather than concentrating mainly on browser control. Its proposed actions can include an intent explaining why a step was selected, while configurable safety policies can classify an operation as allowed, blocked or dependent on confirmation. Optional screenshot scanning is also available to identify potential prompt-injection attempts embedded in an interface.

Google’s July 21, 2026 release materials provide another reason for the change. In the company’s reported OSWorld-Verified comparison, Gemini 3.6 Flash scored 83.0% against 78.4% for Gemini 3.5 Flash, and computer use was presented as a built-in client-side tool rather than a separate specialist model. The same Gemini 3.6 Flash announcement set API pricing at $1.50 per million input tokens and $7.50 per million output tokens.

Those figures should not be read as a direct replay of the web-navigation evaluations used to introduce Gemini 2.5. OSWorld-Verified covers computer interaction under its own harness and conditions, whereas the 2025 announcement discussed several web and mobile benchmarks. The defensible conclusion is that Google now recommends a newer, more broadly integrated model—not that every possible workflow will improve by an identical margin.

How Computer Use works in the current API

A developer still has to build the operating environment around the model. The model sees screenshots and returns proposed actions, but the developer’s application is responsible for moving the pointer, entering text, taking the next screenshot and deciding which sites or resources the agent may reach. Computer Use is therefore a tool for building an agent, not a hosted automation service that independently takes over a user’s machine.

A typical execution cycle contains four parts:

1. The application sends the task, target environment and current screenshot to a supported Gemini model.
2. The model interprets the visible interface and proposes an action such as a click, scroll or keystroke.
3. The client checks the accompanying safety decision and obtains user approval when confirmation is required.
4. The client executes an approved action, captures the resulting screen and continues the loop until completion, failure or termination.

This separation gives developers control over credentials, network access and irreversible operations, but it also creates engineering work. A production implementation needs a secure execution container, an action handler, reliable state capture, task timeouts and logs connecting every screenshot to the action that followed it.

The preview label still sets a hard boundary

Although Gemini 3.6 Flash is the recommended model, Google still describes Computer Use as a preview capability that may produce errors or contain security vulnerabilities. Its documentation advises close supervision for important tasks and recommends avoiding workflows involving critical decisions, sensitive information or serious consequences that cannot be reversed.

The central risk is that an agent treats content visible on a page as an instruction. A malicious message hidden in a document, website or application could try to redirect the model away from the user’s actual request. Detection features can reduce that exposure, but they do not remove the need for sandboxing, restricted permissions and confirmation gates.

For a production deployment, the agent should start with the smallest practical authority. An allowlist can limit navigation to expected domains; a clean container can isolate the session from the host; and sensitive actions such as purchases, submissions or account changes can require explicit approval. Logs should preserve the prompt, screenshots, proposed operations, safety responses and actions actually executed so failures can be reconstructed.

What existing 2.5 integrations should do

An existing integration does not become invalid simply because its model is now labeled legacy. The current documentation continues to list the identifier gemini-2.5-computer-use-preview-10-2025, so teams can assess migration deliberately rather than treating the update as proof of immediate removal.

New projects have a clearer decision: start with Gemini 3.6 Flash unless a tested dependency requires the older model. Existing projects should replay representative tasks against the recommended model and compare completion rate, action count, latency, confirmation behavior and recovery from unexpected interface states. Because the newer models cover more environments and return richer action metadata, migration may also require changes to the client-side handler rather than a simple model-name substitution.

The lasting contribution of Gemini 2.5 Computer Use was the API pattern it established: a model observes an interface, proposes one operation and waits for the application to report the result. That pattern remains central in 2026. What changed is the model developers are expected to place inside the loop—and the scope now extends beyond the browser-focused preview that introduced it.

Also read:

• Meta Muse Image Faces Backlash Over Data Use and Creative Control
• Google AI Overviews Dispute Is Now a Formal EU Antitrust Investigation
• Salesforce Release Management Now Starts in Source Control, Not a Sandbox