Anthropic, OpenAI, and Google Picked Three Different Answers to What Computer Use Should Mean
By mid-2026 the three largest AI labs had committed to genuinely different architectures for computer use, the category of AI agents that operate a computer directly rather than only calling APIs. Anthropic built a portable Computer Use API, a general tool-use model designed to work across different operating environments rather than being tied to one. OpenAI went desktop-native: Codex Background Computer Use shipped April 16, 2026, built macOS-first, with support for running multiple parallel background sessions at once. Google took a third path, adding computer-use style automation directly into Gemini 3.5 Flash in mid-2026, anchored specifically to browser environments rather than full desktop control.
Portable versus native versus anchored
Anthropic's portability bet trades some depth of integration for reach: a general tool-use model that can, in principle, be pointed at whatever environment a developer needs, at the cost of not being purpose built for any single one of them. OpenAI's desktop-native approach goes the other direction, building specifically for macOS and accepting the narrower footprint in exchange for tighter integration with how a real desktop actually behaves, plus the ability to run several sessions in parallel in the background rather than one agent working at a time. Google's browser-anchored choice narrows the scope further still, giving up full desktop control entirely to focus on the single environment, the browser, where most enterprise software and most of the web actually lives.
The security surface each one opens
These are not just product differences, they change what can go wrong. A portable, cross-environment agent has the largest attack surface in principle, since it is designed to operate wherever it is pointed, which means the access controls around it matter more, not less. A desktop-native agent running parallel background sessions on macOS concentrates risk into whatever permissions that machine already has, and multiplies it by however many sessions are running unattended at once. A browser-anchored agent has the narrowest surface of the three, confined to what a browser session can reach, which is also its main limitation: it cannot touch local files or native applications the way the other two can.
Choosing one
For a developer deciding which to build on, the honest tradeoff is portability against depth against exposure, not a single best answer. A team that needs an agent to work across heterogeneous environments and is willing to invest in access controls gets the most from Anthropic's approach. A team building specifically for macOS desktop workflows, and that wants multiple agents running unattended in parallel, is better served by OpenAI's model. A team whose actual task is confined to the browser, filling out web forms, navigating SaaS dashboards, testing web apps, gets the narrowest and probably safest surface from Google's approach, without paying for desktop capabilities it does not need.
None of the three architectures is a strictly dominant choice yet. The right one depends on where the actual task lives, and how much access you're comfortable granting an agent to get there.