Google announced Wednesday that computer use is now a built-in tool inside Gemini 3.5 Flash, available through the Gemini API and the Gemini Enterprise Agent Platform. The feature lets an AI agent see a screen, click, type, and navigate software without human intervention. What makes this release significant for agent developers: it is the same model they already use for function calling, Search grounding, and Maps integration. A single Gemini 3.5 Flash agent can now operate a legacy enterprise application, pull current pricing from Google Search, and interact with a map, all within one model context.
What Changed
Computer use previously required a standalone Gemini 2.5 model launched in October 2025. That model could control screens but could not simultaneously use Search, function calls, or Maps. Developers who wanted both capabilities had to build multi-model pipelines and maintain context across them.
Native integration removes that constraint. According to Google’s computer use documentation, developers now invoke computer use as a single tool parameter alongside any other Gemini tool. The announcement, published by Mateo Quiros, Product Manager at Google DeepMind, positions the launch as a step toward practical enterprise automation, according to TechTimes.
Benchmark Performance and Pricing
On OSWorld-Verified, the standard benchmark for computer-use agents across Ubuntu, Windows, and macOS tasks, Gemini 3.5 Flash scores 78.4. GPT-5.5 sits one position ahead at 78.7. The previous generation, Gemini 3 Flash, scored 65.1, a 13.3-point generational gain, per TechTimes.
Two caveats: Claude Fable 5 leads the OSWorld leaderboard at 85.0%, but was suspended on June 12, 2026 under US export controls. Claude Opus 4.8 holds second at 83.4%. All 16 entries on the leaderboard are self-reported scores with no independent third-party verification as of June 2026.
On cost: Gemini 3.5 Flash charges $1.50 per million input tokens and $9 per million output tokens. GPT-5.5 charges $5 per million input and $30 per million output. For agentic workloads where token volumes compound across many agent turns, that roughly 3x cost differential is material.
How the Perception Loop Works
The model operates through a continuous observe-think-act cycle. A developer’s application captures a screenshot, sends it to the API with a task goal, and the model identifies UI elements (buttons, text fields, menus), reasons about the next action, and returns a structured command: a click at specific coordinates, a keystroke, a scroll, or a form entry. The application executes the command, captures a new screenshot, and repeats until the task completes.
One technical addition distinguishes this from the standalone predecessor: an intent field in each action response. The model now provides a natural-language explanation of why it chose a given action (“Click the search box to type the destination”), making debugging easier for developers working with long-horizon tasks.
Security Layers
Computer-use agents can be hijacked by malicious instructions embedded in the content they encounter, a class of attack known as prompt injection. Google’s response uses three layers: adversarial training during model development to reduce susceptibility, opt-in enterprise safeguards that gate sensitive actions and auto-terminate when injection is detected, and deployment-level guidance recommending sandboxed environments and human-in-the-loop verification for high-stakes tasks.
Google’s own safety documentation recommends against using computer use for critical decisions or sensitive data processing without human oversight.
Production Gaps Remain
The gap between benchmark scores and production reliability is wider in computer use than most AI tasks. The most common failure mode reported by developers is UI drift: page layouts change between screenshots, invalidating the pixel coordinates the model predicted. Dynamic content loading, advertisements shifting layouts, and CAPTCHAs interrupting flows all break the perception-action loop in ways benchmarks do not measure.
Early enterprise customers cited by Google include browser infrastructure platform Browserbase, open-source browser agent framework Browser Use, and enterprise automation platform UiPath.