Computer-use agents have crossed from merely useful to faster than the average human at using a computer, OpenAI says.
Most people still picture an AI agent as something you watch over its shoulder. OpenAI's own computer-use leads increasingly let theirs run unsupervised in the background, doing tasks neither of them wants to do themselves.
"We've gotten computer use now to a place where it's faster than the average human at using a computer. And I think we're getting really close to a place where it is superhuman." — Ari Weinstein
That's Ari Weinstein, who leads OpenAI's consumer computer-use work, speaking alongside James Sun, who leads the enterprise side, on a16z's podcast, hosted by Seema Amble and Fabrizio Serafini.
The full interview is covered here so you can skip it. 45 minutes of audio, 8 minutes of reading.
Here are the 6 takeaways that matter.
a16z · Oct 26, 2026 · 45m · 37m saved
With Ari Weinstein, OpenAI (computer use, consumer) · James Sun, OpenAI (computer use, enterprise)
Key Takeaways
Computer use has crossed from merely useful to faster than the average human, Ari says, and OpenAI is "getting really close" to superhuman
Astra, OpenAI's newest computer-use model, also beats its predecessor at video games "every time," including clearing Fallout in about 30 hours
Enterprises get a dial, not a binary trust decision, via "Guardian," a model that reviews every action a computer-use agent takes
James lets his own AI "dot" act proactively, flagging calendar invites that never made it onto his calendar and monitoring Slack against a loose instruction rather than a rigid keyword rule
Both guests expect legacy software and agent-native software to coexist for a long time, even as WebMCP lets agents and humans share the same site
James's advice to founders: the moat isn't model access, it's speed plus genuine domain expertise — regulatory knowledge, go-to-market motion, how to service a specific customer
For most small businesses, just connecting and deploying the tools that already exist is still unsolved, James says, which is where he sees room for new companies
1. What Astra Actually Changed
Ari said OpenAI's computer-use models were already strong before Astra, and that the new model mainly helps "around the edges": fewer thinking tokens make it a little faster, it's noticeably better at creative tasks, and it closes a reliability gap that's hard to perceive but matters in practice.
Astra's gains are mostly invisible — a gap between 99% and 97% reliability, Ari says
"Our computer use models were already so good before Astra... Astra really helps with a few things around the edges... it's a little bit faster on average because it uses fewer thinking tokens to do the same task." — Ari Weinstein
James added that Astra is OpenAI's most-aligned model to date, scoring better on resisting prompt injection and avoiding unintended actions per its own system card — which he said matters most for enterprise use on sensitive tasks like procurement or vendor communication.
Astra is OpenAI's most-aligned model yet, James says, and it also beats earlier models at video games "every time"
"Astra is also our most aligned model... if you look at the system card for it, it is much better at being resistant to prompt injection [and] not taking any kind of unintended actions." — James Sun
On the lighter side, James said Astra clears Fallout in roughly 30 hours against older models that lose every time.
