How does Gemini 3.5 Flash's computer control performance compare to other models?

On the OSWorld benchmark, Gemini 3.5 Flash scores 78.4, beating Gemini 3 Flash (65.1) and GPT-5.4 mini (72.1). GPT-5.5 scores 78.7 and Anthropic's Opus 4.8 leads at 83.4.

How is Google protecting against misuse of this screen-control feature?

Google uses adversarial training and offers two optional enterprise safeguards: one requires user confirmation for sensitive or irreversible actions, while the other automatically stops tasks when it detects indirect prompt injections. Google also recommends sandboxing, human oversight, and strict access controls.

Where can developers access this feature?

The feature is available through the Gemini API and the Gemini Enterprise Agent Platform. A Browserbase demo and a GitHub reference implementation are also available.

Large Language ModelsTHE DECODERJun 25, 2026Jun 25, 2026, 19:00 JST1 min read

Google's Gemini 3.5 Flash now sees and controls computers

Send on LINE

Key takeaway

Google has integrated computer-control capability directly into its Gemini 3.5 Flash AI model, enabling it to see and operate screens across computers, browsers, and mobile devices.
On benchmark tests, the model ranks highly for this task, making it practical for developers to build automated agents for software testing and office work.
The feature includes built-in safeguards and is available now through Google's Gemini API and Enterprise Agent Platform.

3 Key Points

What happened
Google integrated "Computer Use" into Gemini 3.5 Flash, allowing the model to see, understand, and interact with computers, browsers, and mobile devices on its own. Previously this capability was only available as a separate Gemini 2.5 model. The feature is now available through the Gemini API and the Gemini Enterprise Agent Platform.
Why it matters
On the OSWorld benchmark, Gemini 3.5 Flash scores 78.4, beating Gemini 3 Flash (65.1) and GPT-5.4 mini (72.1), putting it among the top-performing models for computer interaction tasks. This opens the door for developers to build agents that automate software testing, office tasks, and browser workflows across multiple device types.
What to watch
Google has built in two optional enterprise safeguards—one requiring user confirmation for sensitive actions, and another automatically stopping tasks when indirect prompt injections are detected. The company also recommends sandboxing, human oversight, and strict access controls to guard against abuse.

Ask the AI about this article →

Summaries like this, in your inbox every morning.

Get news like this every morning — free

FAQ

How does Gemini 3.5 Flash's computer control performance compare to other models?: On the OSWorld benchmark, Gemini 3.5 Flash scores 78.4, beating Gemini 3 Flash (65.1) and GPT-5.4 mini (72.1). GPT-5.5 scores 78.7 and Anthropic's Opus 4.8 leads at 83.4.
How is Google protecting against misuse of this screen-control feature?: Google uses adversarial training and offers two optional enterprise safeguards: one requires user confirmation for sensitive or irreversible actions, while the other automatically stops tasks when it detects indirect prompt injections. Google also recommends sandboxing, human oversight, and strict access controls.
Where can developers access this feature?: The feature is available through the Gemini API and the Gemini Enterprise Agent Platform. A Browserbase demo and a GitHub reference implementation are also available.

THE DECODERRead Original Article

Send on LINE