Google has unveiled its Gemini 2.5 Computer Use AI, an advanced model capable of performing on-screen actions such as clicking, typing, and scrolling, mimicking human interaction to complete complex digital tasks.
The model, built on Gemini 2.5 Pro, brings powerful visual reasoning and interface navigation abilities, enabling it to operate across web browsers and Android systems. This marks a major leap in AI agent development, designed to automate computer-based workflows efficiently.
According to Google, Gemini 2.5 outperforms rival systems in multiple benchmarks. In the WebVoyager test, it achieved 88.9%, compared to OpenAI’s Computer-Using AI Agent at 87%. It also led the Online-Mind2Web benchmark, surpassing both OpenAI’s and Anthropic’s Claude Sonnet 4.5 models.
| Benchmark Test | Gemini 2.5 Score | OpenAI Agent Score | Claude Sonnet 4.5 |
| WebVoyager | 88.9% | 87% | 85.6% |
| Online-Mind2Web | Top Performer | 2nd Place | 3rd Place |
Google stated,
“Gemini 2.5 Computer Use is a step toward more capable AI agents that can act independently and assist users across digital platforms.”
The company confirmed that Project Mariner and AI Mode in Google Search are already powered by versions of this model. Developers can now access its API via Google AI Studio and Vertex AI, expanding its potential applications across industries.