Gemini 3.5 Flashにコンピュータ操作機能を導入Introducing computer use in Gemini 3.5 Flash

Google DeepMind has introduced computer use capabilities to Gemini 3.5 Flash, allowing the model to operate desktop environments directly and carry out multi-step tasks with limited human guidance. The move extends a capability that has generally been associated with larger, more expensive models to a lighter-weight tier positioned around speed and cost efficiency, which could make agentic automation more practical for high-volume or latency-sensitive workloads.

Computer use refers to a class of AI functionality in which a model perceives a graphical user interface, typically through screenshots or structured accessibility data, and responds by producing actions such as mouse clicks, keyboard input, scrolling, and navigation. Instead of relying solely on a purpose-built API for each application, the model interacts with software much the way a person would. This is significant because a large share of everyday digital work still happens inside interfaces that were designed for humans rather than machines, from web forms and dashboards to legacy internal tools that lack modern integrations.

Technically, computer use systems operate in a loop. The model receives an observation of the current screen state, reasons about the next step toward a stated goal, and emits an action. That action is executed in a controlled environment, a new observation is captured, and the cycle repeats until the task is complete or a stopping condition is reached. Placing this loop on the Flash variant appears aimed at reducing the per-step cost and latency that can accumulate quickly when a task requires dozens of individual interactions. For workflows that involve repetitive navigation, that trade-off between raw capability and efficiency is often decisive.

The announcement fits into a broader industry push toward AI agents that can act rather than only generate text. Anthropic added a computer use feature to its Claude models, and OpenAI has released agentic browsing and task-execution tools under names such as Operator. Google itself has been building toward this with earlier project work and a Gemini 2.5 computer use capability, alongside browser-oriented efforts. Bringing the functionality to Gemini 3.5 Flash suggests Google is trying to move these features from experimental demonstrations toward configurations that developers can deploy at scale.

For developers, the practical appeal is automating processes that previously resisted scripting. Examples that vendors typically cite include filling out multi-page forms, extracting data across several applications, performing quality-assurance checks on web interfaces, and completing routine administrative tasks that span multiple tools. Because the model works from what is on screen, it can, in principle, adapt to interface changes that would break a brittle, hard-coded script, though reliability in messy real-world conditions remains a common limitation for this category of systems.

Those limitations are worth emphasizing. Autonomous interface control raises well-documented safety and security questions. An agent that can click and type can also take unintended or harmful actions, and it may be susceptible to prompt injection, where malicious content embedded in a webpage or document attempts to hijack the model's behavior. Companies deploying computer use features generally recommend running them in sandboxed or permissioned environments, adding human confirmation for sensitive steps such as purchases or data deletion, and constraining the range of actions the model can perform. It is likely that Google is shipping comparable guardrails, given the sensitivity of the capability, though the precise controls will matter as much as the underlying model quality.

Performance on this task is commonly evaluated using benchmarks that measure how often an agent can complete realistic digital tasks end to end. Even leading systems have historically posted success rates well below human levels on the hardest of these evaluations, so the capability should be understood as an evolving assistant for structured work rather than a fully reliable replacement for human operators.

The release also reflects a naming and tiering strategy that has become standard across the field, where a Flash-class model is optimized for throughput and price while a Pro-class model targets maximum reasoning ability. Offering computer use across tiers gives teams flexibility to match cost to the complexity of a given automation. As with earlier Gemini features, access appears likely to arrive through Google's developer platforms such as the Gemini API and its cloud tooling, with adoption depending heavily on how the feature performs on genuine, varied tasks once it is in wider use.