OpenAI achieves superhuman performance in computer use

Summary

The Latent Space podcast covers OpenAI's DevDay 2026 event. Ari Weinstein, leader of the CUA team, argues that Computer Use has reached a '180-degree different' point and can outperform average humans in certain tasks. API product leader Nikunj Handa discusses the Decisions API (a Jev competitor) launched within a week, along with broad agent stack updates including asynchronous tool calling, UltraFast inference, and prompt caching.

Why it matters

Computer Use is the closest current AI capability to a fully digital workforce. OpenAI's claim of superhuman task completion shifts the competitive framework from 'can it do it?' to 'how fast and cheaply can it do it?'. The fact that the Decisions API was completed in a week shows that leading labs' differentiation now relies more on packaging and latency than on raw model breakthroughs. This narrows the competitive advantage of independent agent startups.

Highlights

• GPT-6.1 Sol is offered at one-fifth the cost of Astra and one-seventh the cost for Computer Use workloads. This makes agent loops economically viable at scale.
• The Decisions API is described as an open 'Luna wrapper' launched in about a week. This highlights that OpenAI's advantage now lies in distribution and integration speed rather than a new architecture.
• The Computer History feature in ChatGPT allows the model to learn from everything the user does on their machine. This powerful but privacy-sensitive feature is addressed by the team only as an exit option from the timeline view.

Critical take

The entire content is a self-promotional podcast featuring guests from OpenAI's own team. Therefore, phrases like '180-degree different' and 'superhuman' are unverified marketing language lacking independent benchmarking. The 'superhuman' framing has been softened to 'faster than the average human in some tasks.' This is a much weaker and less specific claim than the title implies.