a16z (Andreessen Horowitz) · Oct 26, 2026 · 44 min
Why OpenAI Is Racing to Make Computer Use Superhuman
OpenAI's Ari Weinstein and James Sun discuss how computer-using agents have advanced with the Astra model, arguing agents are now faster than the average human and close to superhuman. They cover enterprise use cases (ERP, procurement, finance reconciliation), safety controls, the future of software and the open web, startup opportunity, and rapidly growing usage, which they say is only at an early scaling stage.
Where they stand
Computer-using agents are at a scaling moment for both consumers and enterprises, with rapid usage growth and large remaining opportunity for startups and diffusion into slower industries.
High conviction · years
Key takeaways
- Computer use is now faster than the average human and is close to being faster than an expert user, which the speaker calls a sea change for how people interact with computers. 8:48
- Astra uses fewer thinking tokens, is more reliable, and is more resistant to prompt injection, which matters for enterprise trust. 1:39
- Enterprise controls let businesses set human-in-the-loop levels, sandbox apps and websites, and use an auto-approve safety model (Guardian) to audit agent actions. 10:28
- Most promising enterprise use cases are tedious system-of-record workflows: ERP, CRM, procurement, finance reconciliation and sales prospecting, plus creative and design work later. 20:57
- Computer use works with every application without a developer building a connector, unlike MCP servers, which makes it a universal interface for legacy software. 15:28
- Legacy software and new agent-friendly software will coexist; cheap software creation will pressure products to be better, and UI is not going away. 26:25
- Agent use on the open web raises unresolved questions for ad-based business models, while standards like WebMCP let humans and agents share sites. 12:07
- Usage is growing very fast ('YC level of growth week over week') yet remains early; the next inflection is cloud-based agents reaching phone-first consumers. 41:51
Calls and positions
- Superhuman computer use will change how everyone interacts with computers, as agents would default to doing anything done manually faster.
“I think it's going to change the way we all interact with computers”
- Legacy software will not disappear overnight, and computer use will put pressure on companies to make software work better with agents.
“I don't think that that stuff's going to go away overnight”
- The industry is at the scaling moment of computer use for both consumer and enterprise.
“we're at the scaling moment of computer use both for consumer and enterprise”
Views expressed
- Computer use is already faster than the average human and is approaching superhuman, faster than an expert computer user.
“we've gotten computer use now to a place where it's faster than the average human at using a computer”
- OpenAI believes it has the fastest, cheapest and most reliable computer use product.
“we have the fastest, cheapest and most reliable computer use product”
- Controls and auditability become more important as computer-using agents become superhuman and can no longer be fully observed.
“the control part becomes more and more important as our computing agents become also superhuman”
- Computer-using agents with Astra can do anything a person can do on a computer.
- UI is not going away; cheap software will force software to be better rather than everything becoming MCP and API.
“I don't think UI is going away”
- Startups still have a large greenfield because AI diffusion into vertical industries is slow and labs are stretched thin.
“I'm still very bullish on startups”
- Agent usage is growing extremely fast but is still early, with only a portion of users using computer use.
“it's still so early”
- The agentic browser product saw weak adoption initially but now sees very fast week-over-week growth after being built into existing products.
“it's like YC level of growth week over week”
- Ad-based publishers face a business model conflict because agent impressions do not count as real ad revenue.
“most places a agent's ad impression does not”
Companies, projects and topics mentioned
More from a16z (Andreessen Horowitz)
- Amazon CEO Andy Jassy on the $220 Billion AI BuildoutOct 23, 2026
- The Frontier Ideas a16z Wants to Fund Now | Marc Andreessen, Ben Horowitz & Chris DixonOct 22, 2026
- OpenAI’s Head of Health AI on Whether You Should Trust ChatGPTOct 20, 2026
- Why a $1M Startup Wasn’t Worth Building in the Age of AIOct 19, 2026
- What Comes After Stablecoins? (ft. Circle CEO Jeremy Allaire and Chris Dixon)Oct 16, 2026
- Why AI Can Write Code But Can’t Keep It Running (Yet)Oct 16, 2026
Summary generated by AI from the video's captionsand checked against it; quotes are verbatim. Processed Oct 27, 2026.
Read on Risk summarises what other people said in public videos and podcasts. Summaries are generated by AI and checked by code against the transcript, but can still be wrong; follow the source links before relying on anything. This is not investment advice and nothing here is a recommendation to buy or sell. Disclaimer ·How this works