OpenAI Says Computer Use Is Now Faster Than the Average Human and Nearing Superhuman, With Agent Usage Growing at "YC-Level" Rates
October 26, 2026 – a16z podcast interview with OpenAI's Ari Weinstein and James Sun on computer-using agents
OpenAI executives made two claims in an a16z podcast interview. First, the company's computer-using agents have crossed the speed of the average human operator and are approaching expert-level speed. Second, usage has accelerated sharply in recent months after a long period of limited adoption. Ari Weinstein, who works on computer use at OpenAI and leans toward consumer products, said: "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, meaning it is faster than an expert computer user." He called that threshold a potential "sea change" because "anything that you could do manually, it could do faster." The claims are self-reported and came without benchmarks or usage figures. Even so, the interview is a useful look at how the team sees its product, its adoption curve and the obstacles that remain.
Usage Has Inflected, but Penetration Is Still Low
Weinstein described a sharp change in trajectory. The team began building computer-use products when "very few people were using them." The underlying models then "all of a sudden" became strong at these tasks earlier this year, and in the past couple of months "the product's growing like crazy." He was candid about the limits: only "some portion" of OpenAI's user base uses computer use today.
Sun, who joined OpenAI in January to work on the Atlas browser, offered a more candid account of the earlier struggle. The agentic experience in Atlas "wasn't actually a big selling point" because most people treated it as a normal browser. Sun's "RIP Atlas" aside suggests the standalone browser has since been folded into broader products. Once model capabilities improved, OpenAI built browser use into existing products that already had agentic usage, and Sun said the company has seen "YC-level of growth week over week."
Sun identified the next inflection as mobile and cloud. Desktop use remains "very much a power user thing," while most consumers live on their phones. The aim is to give every chat user "a powerful computer that they don't even need to understand how to use," which Sun framed as the path to a billion users. Both executives said they believe the product is at "the scaling moment" for consumer and enterprise simultaneously.
What the Astra Model Changes
The interview centered on Astra, OpenAI's latest model. Weinstein was measured about its impact on computer use, noting that the company's computer-use models "were already so good before Astra." In his experience, Astra improves the product "around the edges" in three ways. It is faster on average because it uses fewer thinking tokens for the same task. It is better at creative work, as in a launch demo where it drew in Microsoft Paint on voice instruction. And it is more reliable, which he framed as the difference between 99% and 97% reliability: "hard to perceive but really important."
Sun added that Astra is OpenAI's "most aligned model," citing the system card for much better resistance to prompt injection and fewer unintended actions. For enterprise work involving procurement or communication with vendors and customers, he argued, that makes deployment feel considerably safer. Sun also said internal testing had Astra beat the previous 5.6 model at video games every time, and cited public posts of it completing Fallout in roughly 30 hours. He conceded he was not sure what that translates to commercially.
Weinstein described OpenAI's position as having "the fastest, cheapest and most reliable computer use product." That is an assertion, not a demonstrated result. He also said the company will keep offering models at multiple quality and price tiers, and that cheaper models are improving at computer use. He expects more variation over time in which models are used for which tasks.
Enterprise Controls: Sandboxes, Approvals and an Automated Reviewer
As agents get faster than the humans overseeing them, Sun said, the control problem becomes central: "Once you can get to a point where you can no longer observe and audit everything it's doing, how do we give especially enterprises and businesses the ability to control the amount of human in the loop there is, and then the auditability?"
OpenAI has launched enterprise controls built on two layers. The first is deterministic. Administrators can sandbox the agent to approved applications and websites, and can require human approval for specific actions such as downloads, uploads or visits to new sites. The second is a model-based reviewer the company calls Guardian, or auto-approve, which evaluates each action the agent takes for safety and alignment. Consumers may prefer to approve nothing, while enterprises can decide which action types proceed automatically.
Sun said enterprise hesitation often comes down to accountability and a feeling of safety, not capability, and that the deterministic layer "gives enterprises a lot of feeling of safety" even though the model is better at resisting injection than most people. OpenAI also presents internal and third-party alignment evaluations to administrators. On the consumer side, Weinstein said design choices such as a visible, animated cursor and permission prompts for new applications are meant to build trust and ensure users do not fear "runaway stuff" on their machines.
Where Enterprises Are Deploying Agents
Sun said the enterprise case is "fairly obvious" because large companies run legacy software, some 15 to 20 years old, that has no MCP tools and is awkward even for human users. Some companies have built "skills" to teach the model to operate those systems. He sorted use cases into two buckets. The first is what he called boring but valuable automation across systems of record such as ERP and CRM. Examples he cited were procurement, where annual equipment orders currently pass from team spreadsheets to an ops team, then into procurement software and then out to vendors. Month-end finance reconciliation across multiple ledgers lacking APIs was another, as was sales prospecting that feeds data into CRM and outreach tools. Market research also fits the category.
The second bucket is more ambitious creative and technical work enabled by improved visual capability, including circuit design, architecture, hardware engineering and game design. Sun said enterprises are not yet thinking about these, since they are focused on the simpler opportunities, but he expects usage there as customers learn what the model can do.
Weinstein pointed to consumer analogues, including filling out DMV forms, unsubscribing from email lists, and maintaining personal finance and meal-planning spreadsheets. He also highlighted voice control of the whole computer as the category he is "most excited about."
Dot and the Move Toward Proactive Agents
The executives spent considerable time on Dot, an OpenAI product in which each agent gets its own cloud computer with a full web browser, able to buy items and book appointments. Sun said its main differentiator is proactivity. It flagged external meeting invites that had not reached his calendar and monitors Slack for critical feedback from his reporting chain using high-level instructions. It can also drive both his local browser and its cloud machine to pull Sentry logs, analyze them and report back. He argued Dot works best for "proactive and long-running tasks where they're not the same shape every single time."
Weinstein said the next step is real-time collaboration between user and agent. Today's product is heavily background-oriented, which he called a strength but a limit. He wants agents that "adapt to you instead of you adapting to it."
Open-Web Economics Remain Unresolved
Sun was direct about a commercial problem on the open web. Within company-approved software, agent automation is welcome. On the internet, advertisers and ad-funded publishers have business models "not in line with having agents use our product," and an agent's ad impression "does not actually count towards real revenue." He said OpenAI is working through the question with its partnerships team and offered no resolution.
For sites that want to accommodate agents, he pointed to WebMCP, a standard OpenAI supports in its app browser. It lets people and agents use a site together, with the agent handling tedious steps and the human taking over for browsing or checkout. Sun said the model keeps the site itself as the interface rather than reducing it to an API.
Software Form Factor and Startup Opportunity
Asked whether legacy software or AI-native applications will win, Weinstein said both will persist. Legacy systems written 10 to 40 years ago will not disappear quickly, and computer use "may put more pressure on companies to make software work better with agents," because all software will work with agents by default. Sun argued that cheap software creation will force quality upward, rejecting a future in which the internet becomes "faceless" agent-to-agent traffic. He described building a personalized training tool for his wife and predicted software designed "for a hundred people or one person." In B2B, he said, users will still need an interface for audit logs and control, and "why should we settle for bad software?"
Both executives, former founders, remained constructive on startups despite OpenAI's breadth. Sun said OpenAI's teams are still "stretched thin" relative to the problems. He cited a vertical example in which an agent photographs restaurant storage, forecasts needs from projected traffic, orders groceries through procurement software and adjusts staffing in the ERP. He said many small and mid-sized businesses struggle with understanding, connecting, deploying and servicing AI tools, which leaves room for specialists who understand regulation, go-to-market and customer service in specific industries.
Product Coherence Is the Near-Term Problem
Sun acknowledged an uncomfortable product reality. Users currently face a Chrome extension using their own passwords, an in-app desktop browser with separate cookies and logins, a cloud browser in ChatGPT, and another cloud computer in Dot, none of which share state. "Why do I have these four different browsers that all don't talk to each other?" he said the average consumer will ask. He called computer use "the most easy to understand AI product" in concept, but said implementation complexity leaks through to users. Making the experience coherent and consumer-grade is his priority for next quarter, and he described a "shared intelligence layer" so that agents and applications can exchange context instead of fragmenting into separate website-specific agents.