Breaking

Google brings native computer use to Gemini 3.5 Flash

June 25, 2026 at 12:22 EDT

AI Agents
Foundation Models

Google DeepMind announced on June 24, 2026 that it has built computer use into its lightweight Gemini 3.5 Flash model as a native tool. Developers can now build agents that see and operate browser, mobile and desktop interfaces within a single model, without calling a separate dedicated model.

June 24 · Google DeepMind

Gemini 3.5 Flash Now Sees and Operates the Screen Natively

"Computer Use" is now a built-in tool of the lightweight Flash model — clicking, typing and acting across browser, mobile and desktop without a dedicated model or external wrapper, at Flash's speed and low cost.

OSWorld-Verified — agentic GUI tasks in real environments

78.7%

GPT-5.5

78.4%

Gemini 3.5 Flash

72.5%

Claude Sonnet 4.6

65.1%

Gemini 3 Flash
(prior gen)

Up sharply from the prior generation — now neck-and-neck with the frontier leader.

input tokens supported (64K output)

environments controlled: desktop, mobile, browser

Free

Computer Use tool calls during Early Access

Cua-Bench · KiCad CAD suite

25 electrical-engineering tasks · 200-step budget · mean reward

0.267

Gemini 3.5 Flash
5 full + 3 partial solves · highest reward

lower

GPT-5.5
6 full solves, no partials · trailed on reward

Praised

Pixel-accurate grounding on zoomed-in targets
Analog-design reasoning — circuit values, datasheet checks
Practical desktop control at Flash's speed and cost
Latency for continuous automation has fallen substantially

Limitations

"Hallucination of screen state" after repeated screenshots
Acts on nonexistent UI — gap between observation and action
Complex tasks time out within the 200-step budget
No model yet solves the KiCad suite outright

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.