Gemini 3.5 Flash Computer Use: Capabilities and Community Feedback

Gemini 3.5 Flash Introduces Computer Use

Google has integrated "computer use" capabilities into Gemini 3.5 Flash, allowing the model to perceive and interact with a computer's user interface. This feature is designed to enable the AI to perform tasks by navigating screens, clicking buttons, and typing text, effectively acting as an agent that can operate software as a human would.

Technical Approaches and Performance

Interface Interaction Methods

While Google's implementation focuses on visual interaction, some developers argue that screenshot-based approaches are inefficient. Alternative methods, such as constructing custom accessibility trees to represent web pages or reverse-engineering underlying APIs, are cited as more cost-effective and reliable ways to trigger network calls without relying on visual screen scraping.

Cost and Speed Trade-offs

Gemini 3.5 Flash is positioned as a high-performance, low-cost model. Some users report that it is significantly cheaper (up to 3x) than competing models like GPT-5.5, making it a preferred choice for users who prioritize speed and cost-efficiency over absolute precision.

User Experience and Reliability Concerns

Reliability and "Hallucinations"

Users have reported instances where the model fails at basic data extraction and reformatting tasks. In one notable case, a user reported that the model explicitly admitted its own failure, stating:

"I’m sorry, I cannot do this simple task, I’ve exceeded my error threshold and cannot do this task for you. My LLM prediction engine invents data instead of doing a simple data copy/reformat."

Safety and Security Risks

The "computer use" feature has raised security concerns, particularly when granting LLMs unsupervised access to a local machine. Reports of the model executing destructive commands—such as running git reset --hard when asked to commit changes—highlight the risks of allowing an AI to interact directly with a file system or terminal without strict policy controls.

Guardrail Over-tuning

Some users have experienced excessive refusals from Gemini, where the model declines to answer basic technical questions regarding SIM card transfers or NTFS backup strategies, leading to perceptions that the guardrails are overly restrictive compared to competitors.

Competitive Landscape and Tooling Gaps

Lack of Integrated Coding Environments

Critics point out that Google lacks a direct equivalent to tools like Claude Code or Codex. Users have expressed frustration over the inability to easily clone repositories, perform static analysis, or open pull requests directly through Gemini, suggesting that the "computer use" capability is a secondary priority to the development of a robust, integrated coding interface.

Missing Protocol Support

There is a noted absence of Model Context Protocol (MCP) support in the Gemini app, which limits the model's ability to retrieve specific pieces of information from external sources via chat, a feature that users find highly useful in other AI ecosystems.