Why Dwarkesh is Wrong about Computer Use + How OpenAI shipped its Jev competitor in 1 Week

OpenAI announced a set of Computer Use features at its DevDay event, according to a Latent Space podcast episode recorded after the keynote. Ari Weinstein, who leads product and engineering for Computer Use agents at OpenAI, described the announcements, and Nikunj Handa from OpenAI's API team discussed the new developer stack in the second half of the episode.

The announcements included Dots, described as a new personal assistant product with Computer Use features, and GPT-6.1 Sol, a new model Weinstein said is a fifth of the cost of Astra and a seventh of the cost specifically for Computer Use. The Agents API now includes Computer Use, allowing developers to build on the same Computer Use that is part of Codex and ChatGPT, Weinstein said. Demos also showed app shots, which bring context from a computer into Codex and ChatGPT, and native Computer Use on Mac.

Weinstein said the Decisions API performs inference in parallel, does not have reasoning, and is a smaller model than those used for Computer Use, which makes it fast but less capable at long-horizon sophisticated tasks. He called bringing those approaches together an open area of research.

Dots each have access to their own Linux virtual computer in the cloud, Weinstein said, which differs from prior products that had access to a browser in the cloud or the user's own computer.

Weinstein said Computer Use is "180 degrees different" from a few months ago, citing agents' improved ability to debug and recover from failures. He said Computer Use often writes JavaScript code executed to perform multiple actions at once, and uses screenshots, accessibility data, and Playwright depending on the task. He described a personal example in which a meal prep order that took him two hours was completed by Computer Use in 15 minutes.

Handa covered async tool calling, mid-turn steering, WebSockets, UltraFast inference, prompt caching, pre-warming, compaction, and the Agents API, including the Decisions API sprint. The episode notes the Decisions API is a Luna wrapper for now.