Google’s October 30 release notes mark Computer Use generally available across the Interactions and GenerateContent APIs. For developers evaluating the capability, the central distinction is who acts: the model proposes interface actions, while the application executes them. General availability establishes release status; it does not establish measured reliability or remove the engineering work needed to operate the loop.
What the model returns—and who clicks
According to Google’s Computer Use documentation, an application sends a prompt and screenshot. The model returns suggested UI actions as function calls. The application executes those actions, captures another screenshot, and sends the updated view back to continue the process.
That makes the client application an active participant throughout the workflow. It supplies the screen, handles execution, and returns the next observation. The practical implication is that Computer Use provides a model-driven action capability, while developers still own the surrounding automation system and its stop conditions.
The execution environment remains your responsibility
Google describes browser, mobile, and desktop environments for Gemini 4 and Gemini 3.x. Its guide directs developers to provide a client-side handler and a sandboxed virtual machine or container. Those requirements matter when interpreting the GA announcement: an available API feature still needs an environment in which its proposed actions can run.
The release note does not establish identical behavior, model eligibility, regional availability, or account-tier access across both API surfaces. Developers should therefore distinguish the broad release announcement from the specific configuration they intend to use.
This execution boundary also provides a useful comparison with Microsoft’s cross-product Copilot harness: here, the defining mechanism is the screenshot-and-action loop, with the client carrying out each proposed move.
Safety signals still need application policy
Google documents intents, configurable safety policies, and opt-in screenshot prompt-injection detection. A response may include a safety decision indicating that an action is allowed, requires confirmation, or is blocked. These are documented mechanisms, not independent evidence of how reliably a workflow will avoid harmful actions.
The reviewed sources provide no miss rate or false-positive rate for prompt-injection detection. An allowed decision should therefore inform application policy rather than serve as proof that every permitted action is safe to execute.
Google warns of errors and security vulnerabilities, advising against critical decisions, sensitive data, or actions where serious errors cannot be corrected. Its general safety guidance assigns application-level safety responsibility to developers and recommends risk assessment, testing, monitoring, and iterative mitigation. Our coverage of OpenAI’s agent monitoring addresses a related question; monitoring for misalignment and controlling interface actions are distinct concerns.
API choice affects the data-handling review
Google’s Interactions overview says that API has been generally available since June 2026 and is recommended for new projects. GenerateContent remains supported but is described as legacy; Google says new models, tools, and agent features will launch on Interactions.
The overview also says Interactions requests are stored by default, with store=false available for stateless behavior. The reviewed pages do not establish screenshot-specific retention duration or equivalent data handling on GenerateContent. Developers should review their chosen surface’s settings before sending screenshots.
Token prices do not establish task cost
The pricing page lists per-token rates, including separate input and output rates for Gemini 3.8 Flash. It does not quote a typical completed Computer Use task cost. Estimating that requires request counts and token usage across the screenshot-action loop.
A bounded pilot can answer the remaining questions
A reasonable next step is a sandboxed pilot with reversible tasks, tracking action accuracy, confirmations, blocked actions, recovery behavior, and token use. This is proposed evaluation guidance, not testing performed for this report. No API calls, authenticated account checks, or independent performance evaluation informed this article. GA confirms release status; workload-specific reliability, latency, and safe completion remain questions requiring evidence.