OpenAI's Weinstein: Computer Use Now 180 Degrees Different

OpenAI's DevDay announcements arrive against a backdrop of skepticism. Three months ago, Dwarkesh Patel publicly asked why progress on computer use had been so slow, given that it is so clearly verifiable — a question that, in the podcasters' framing, upset many in the field. Ari Weinstein, who worked on computer automation at Apple and then founded Sky before joining OpenAI, is positioned as the direct answer to that skepticism.

The changes Weinstein describes are not just better models. He points to a shift in how Computer Use agents operate: instead of screenshotting, scrolling, and repeating, agents now write and execute JavaScript to perform many actions at once, and draw on screenshots, accessibility data, the DOM, and Playwright depending on the task. The features demoed at DevDay — Dots with their own Linux virtual computers, app shots that capture rich metadata rather than flat screenshots, and GPT-6.1 Sol at a seventh of the cost for Computer Use — are the packaging of that shift for developers.

The stakes hinge on whether these gains transfer from demo conditions to complicated, long-horizon tasks. Weinstein himself flags that combining the fast Decisions API with the larger Computer Use models remains an open research question, and he notes that in some benchmarks a non-trivial share of time is now spent simply waiting for third-party sites like doordash.com to load. Whether Computer Use reaches what he calls "literally superhuman" performance — or gets stuck on those paper cuts — is the open question developers building on the new APIs will be watching.