BREAKING
Gemini 3.5 Flash Gets Computer Use
From Preview Model to Built-in Tool
Before
2.5 preview
●
Separate computer-use model
●
Extra setup needed
Now
3.5 Flash
●
Built-in tool
●
No special setup
How the Visual Agent Loop Works
1
Analyze screenshot
↓
2
Return click(x,y,intent)
↓
3
Client executes
↓
4
Send new screenshot
78.4% on OSWorld-Verified
3.5 Flash
78.4
3 Flash
65.1
0
%
Toolathlon
0
%
Finance Agent v2
0
Environments
Easy to Try, Sandbox a Must
AI NEWS BLITZ
Google has built screen-viewing computer control directly into Gemini 3.5 Flash.