On June 24, 2026, Google DeepMind announced that computer use is now a built-in tool in Gemini 3.5 Flash. The capability previously existed only as a standalone Gemini 2.5 computer use model; it now ships natively in the mainline Flash model, alongside function calling, Search grounding, and Maps grounding. Developers can combine these tools within a single model to build agents that can “see, reason and take action across browser, mobile and desktop environments.”
The announcement, posted by product manager Mateo Quiros, targets long-horizon and enterprise automation tasks such as continuous software testing and knowledge work across professional applications. Google demonstrated two examples: the model analyzing the Gemini app to return a categorized feature list, and auditing its own documentation for accessibility issues.
From Standalone Model to Built-In Tool
The headline here is not just a new capability but where that capability now lives. Doing computer use used to mean calling out to a dedicated older model and stitching the result back into your main workflow. Now Gemini 3.5 Flash itself can operate a graphical interface, so an agent can decide within the same reasoning pass whether to call an API, search the web, or simply click through the screen.
That is a real simplification for agent architectures. Computer use shines precisely where no API exists: legacy systems and desktop software. Making it a built-in tool of a Flash-tier model closes the last mile of interface automation inside Google’s agent toolchain.
Access is open now: developers and enterprises can use it via the Gemini API and the Gemini Enterprise Agent Platform, with a hosted interactive demo environment and a reference implementation available on GitHub.
OSWorld: Where 78.4 Lands
The Decoder’s coverage cites the OSWorld benchmark: Gemini 3.5 Flash scores 78.4. For context, that beats Gemini 3 Flash at 65.1, GPT-5.4 mini at 72.1, and Gemini 3.1 Pro at 76.2; it ties Anthropic’s Claude Sonnet 4.6 at 78.4 and sits a single step behind GPT-5.5 at 78.7. The current leader is Anthropic’s Opus 4.8 at 83.4.
Put together, the picture is clear: computer use has entered a phase where the front of the pack is biting at each other within margin-of-error distances, and nobody holds a runaway lead. Google calls this its “best performance yet for agentic computer use tasks,” though no independent third-party verification was provided.
Safety: Adversarial Training and Enterprise Guardrails
The biggest risk for agents that operate real interfaces is prompt injection: instructions hidden in a malicious webpage or document can hijack an agent into doing something unintended. Google’s first line of defense is targeted adversarial training, reducing the odds that the model gets misled by injection attacks in live environments.
The second line is two optional enterprise safeguards: one requiring explicit user confirmation before sensitive or irreversible actions, and another that automatically halts a task when an indirect prompt injection is detected. Google also recommends defense-in-depth — secure sandboxing, human-in-the-loop verification, and strict access controls, all three together.
What It Means for Agent Developers
Three direct effects. First, selection logic changes: when a Flash-tier model ties frontier models on OSWorld, “should I pay frontier prices for computer use” becomes a calculation worth redoing. Second, safe defaults become a differentiator: confirmation flows and injection halts are now checkbox options, which moves the security terms enterprise buyers care about into the platform layer. Third, computer use is shifting from demo technology to daily engineering — document review, accessibility checks, and continuous testing are the unglamorous but high-value processes where it lands first. Prompt injection risk still needs your own assessment; sandboxing and human oversight remain non-negotiable.
Sources
- Introducing computer use in Gemini 3.5 Flash — Google
- Google bakes computer control directly into Gemini 3.5 Flash — The Decoder
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.