Gemini 3.5 Flash gets built-in computer use

Gemini’s computer-use story is moving from a specialist preview model into the main Flash lane. Google says computer use is now a built-in tool in Gemini 3.5 Flash, letting developers build agents that can look at a screen, reason about the next step, and return UI actions such as clicking, typing, scrolling or using keyboard shortcuts across browser, mobile and desktop environments.

That matters because computer use is one of the places where agent platforms stop being demos and start touching real workflows. Instead of routing a UI task through a separate Gemini 2.5 Computer Use model, teams can keep the interaction inside gemini-3.5-flash and wire it through the Gemini API or Gemini Enterprise Agent Platform. The obvious use cases are not chatbots; they are software testing, repetitive form work, web-app operations, and enterprise knowledge tasks where no clean API exists.

Google’s product angle is also defensive in a useful way. OpenAI, Anthropic and browser-automation startups have all pushed agents that operate software directly, but the hard part for enterprise buyers is not just model capability; it is controlling what an agent is allowed to do after it sees untrusted web content. Google is pairing the Flash integration with targeted adversarial training, explicit confirmation for risky or irreversible actions, and an optional prompt-injection detection setting for screenshots. Those safeguards do not remove the need for sandboxing or human review, but they make computer use feel more like a governable platform feature than a lab model bolted onto a browser.

The implementation is still a developer loop, not magic remote control. The Google Cloud documentation says apps send the model a user goal plus a screenshot, receive function_call actions with normalized coordinates or text, execute those actions in a client environment such as Playwright, then send back a fresh screenshot and URL. During preview, the same docs say the tool is supported for gemini-3.5-flash and gemini-3-flash-preview, assumes the Python Gen AI SDK, and is not supported in other SDK languages or the Google Cloud console.

Builders can try the feature through the Gemini API docs, the Gemini Enterprise Agent Platform, Google’s computer-use preview reference implementation, or a Browserbase-hosted Gemini Browser demo. A practical first test is a browser-only workflow with low business risk: open a staging app, ask the agent to find a broken checkout path or accessibility issue, require confirmation before submission, and log every proposed action before execution.

The limitations are important enough to plan around. Google labels the Computer Use model and tool as pre-GA in Cloud docs and warns that actions can be wrong or unsafe, especially when adversarial instructions appear on screen. Pricing, quotas, supported regions, and enterprise entitlement details were not disclosed in the launch post, so production pilots should treat the release as an API/platform preview rather than a fully governed automation replacement.

References:

• Google Cloud Computer Use model and tool docs
• Previous Gemini 2.5 Computer Use model launch
• Reference implementation on GitHub
• Browserbase Gemini Browser demo

from google import genai
from google.genai.types import (
    ComputerUse, Content, Environment, GenerateContentConfig, Part, Tool,
)

client = genai.Client()

response = client.models.generate_content(
    model="gemini-3.5-flash",
    contents=[
        Content(
            role="user",
            parts=[Part(text="Open our staging app and find the broken checkout step.")],
        )
    ],
    config=GenerateContentConfig(
        tools=[
            Tool(
                computer_use=ComputerUse(
                    environment=Environment.ENVIRONMENT_BROWSER,
                    enable_prompt_injection_detection=True,
                )
            )
        ]
    ),
)

# Your app still has to inspect the returned candidate content parts,
# execute proposed function calls in Playwright or another sandbox,
# require user confirmation for risky actions, and send back screenshots.
first_candidate = next(iter(response.candidates))
print(first_candidate.content.parts)