Google expands Gemini 3.5 Flash with “computer use” capability

Alphabet’s Google has introduced a new capability in its Gemini 3.5 Flash model: a built-in “computer use” tool designed to let the AI system interact with a computer interface as part of how it completes requests, according to a post on Google’s official blog.

In the announcement, Google frames “computer use” as an extension of Gemini’s ability to execute practical work, not just generate text. The company describes the feature as being integrated directly into Gemini 3.5 Flash, rather than requiring separate tooling or manual steps from the user.

Google also positions the update as a step in making AI more agent-like, with the system able to follow instructions in the context of what appears on a user’s screen. The implication is that the model can translate a user goal into a sequence of actions, such as navigating an interface and using controls, while staying within the constraints Google sets for safe operation.

While the blog post focuses on the product introduction, it does not, in the information available here, spell out quantitative performance results, benchmarks, or detailed limitations of the tool, such as accuracy rates, latency ranges, or failure modes. It also does not provide details on whether the feature is available to all users by default, whether it is limited to certain regions, or how user permissions are handled beyond the conceptual description.

For Alphabet, the move fits into a broader industry shift toward “AI agents” that can take action in software environments. Computer-use features are a key differentiator because they aim to reduce the gap between natural-language requests and the concrete steps required inside apps and browser-based workflows.

The company’s disclosure suggests that Gemini 3.5 Flash is being treated as an important product tier. “Flash” typically indicates an emphasis on faster and more efficient model behavior, which would matter for scenarios where users want the system to carry out tasks interactively rather than waiting for slower, purely text-based responses.

Still, several practical questions remain unanswered in the public-facing announcement as reflected in the material available here. Google does not provide the specific interfaces supported, the scope of actions the system is allowed to perform, or the extent of human oversight for higher-risk tasks within the computer-use flow.

Looking ahead, investors and users will likely watch for follow-on information about rollout timelines, supported use cases, and any safety guardrails or evaluation results Google may publish. Additional clarity on accessibility, supported platforms, and measurable reliability would also be important for assessing how broadly the feature could be deployed.