Latent Space: The AI Engineer Podcast

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

39 min

This episode of Latent Space covers two major OpenAI announcements from Dev Day: the release of the computer use API in the Agents API and the new GPT-6.1 model, alongside the introduction of the Decision model. Ari, who leads computer use agents, discusses technical improvements, measurement approaches, and the future of superhuman computer use, while Nikunj, product lead for the API team, explains the new async function calling, structured outputs, and the decision model's low-latency capabilities. The conversation also touches on developer adoption, caching, and the evolution of AI-native API design.

Summarized by Podsumo

🎧 Listen 🎙️ Ask about this episode

✨ Key Takeaways

1

OpenAI announced GPT-6.1, which is 5x cheaper than Astra overall and 7x cheaper for computer use specifically, enabling more accessible agentic applications.
2

The new Agents API now includes computer use, allowing developers to build on the same technology powering Codex and ChatGPT, with features like app shots and native Mac integration.
3

Computer use is now faster than the average human at many tasks, and the next frontier is achieving superhuman speed and reliability, which will enable real-time, agent-first workflows.
4

The Decision model, part of GPT-6, offers sub-200ms latency for fast classification and structured outputs, though it lacks reasoning, making it ideal for simple, high-volume tasks.
5

OpenAI emphasizes iterative deployment, encouraging developers to use the new Responses API and provide feedback to shape future improvements, especially around performance and caching.

💬 Notable Quotes

"_"We're all still in the process of getting comfortable with this technology... it's important to think about."_ — Ari"
"_"The next frontier is to have computer use be literally superhuman... I think we'll start to default to doing certain things in agents that we've become accustomed to doing manually._ — Ari"
"_"You're building an AI cloud... you have to do the AWS invention, but you're doing the AI native versions of each of these._ — Nikunj"