Why OpenAI Is Racing to Make Computer Use Superhuman
Ari WeinsteinJames SunSeema AmbleFabrizio Serafini
- Ari Weinstein’s speed thesis is that expert-level computer use could make delegation the default. He says agents are already faster than the average human at using a computer and are “getting really close” to exceeding experts. The adoption threshold is overcoming the feeling that “I could’ve just done that myself,” particularly when a task must finish before other work can happen.
- Computer use gives agents access to legacy software without waiting for new integrations. Weinstein’s framing is that “computer use works with everything,” using applications through the same interface as a person. He expects decades-old systems and cheaper, newly built software to coexist, creating room for products that wrap legacy workflows in better experiences.
- Enterprise adoption depends on enforceable boundaries alongside better models. James Sun describes application and website restrictions, configurable approval requirements, and a system he calls Guardian—or “auto-approve” in the app, he thinks—that checks actions against safety and the user’s intent. As agents become too fast to watch continuously, “the control part becomes more and more important.”
- The clearest immediate workflows connect tedious work across systems of record. Sun’s procurement example moves 1,000 spreadsheet rows through purchasing software and vendor outreach; finance reconciliation offers another opening where APIs are missing. Seema Amble’s interjection captures the opportunity: “boring is opportunity.”
- Vertical expertise and deployment remain openings for startups despite broader lab capabilities. Sun argues that understanding regulation, go-to-market and customer service still matters, while OpenAI’s own teams remain stretched. His “speed and pain” moat is illustrated by a restaurant workflow connecting inventory photos, demand forecasting, purchasing and employee scheduling.
- Agents create pressure on web monetization without eliminating the value of interfaces. Sun thinks agent ad impressions usually do not generate revenue, while WebMCP could let people browse and agents handle tedious tasks. His broader position is explicit: “I don’t think UI is going away”; Fabrizio Serafini adds that enterprise users still need payment explanations, audit trails and control.
- Cloud distribution could broaden adoption, but product coherence remains unfinished. Sun reports “YC-level growth week over week” after browser capabilities entered products already used agentically, without providing figures. Reaching a billion users is an ambition tied to giving phone users cloud computers; fragmented browsers, credentials and context remain obstacles, while Weinstein wants agents to “adapt to you instead of you adapting to it.”
1. Superhuman speed would make delegation the default
Weinstein says computer use is already faster than the average human at using a computer and “getting really close” to being superhuman, meaning faster than an expert computer user. Background jobs can take time, but watched tasks—or tasks that must finish before something else can happen—must overcome “I could’ve just done that myself.” Once agents are faster, he expects delegation to become the default unless someone specifically wants to perform the task.
Weinstein says Sol was already strong at computer use. Astra is “a little bit faster on average” because it needs fewer thinking tokens for the same task, and feels more reliable around the edges. His comparison between 99% and 97% reliability illustrates why a small improvement matters; he presents that assessment as anecdotal experience.
Creative gains appear in the Microsoft Paint launch demo Weinstein cites. Sun adds that Astra beats 5.6 “every time” in internal game matchups and cites tweets describing Astra beating Fallout in “about 30 hours or something like that.” His bridge to practical work remains loose: “I know that probably translates to something.”
Weinstein calls OpenAI’s offering the “fastest, cheapest, and most reliable” computer-use product, without supplying comparative figures. He expects improving cheaper models to produce more variation in model choice by task. Sun still sometimes handles a two-button task himself because that is easier; further speed gains could remove even that boundary.
2. Enterprise readiness requires controls beyond model capability
Amble’s pushback is accountability: employers can hold a trained person responsible, even if that person performs worse. Sun generally finds Astra capable of “anything that a person can do” on a computer, with sensitive actions and messages that should feel personal warranting involvement. Repeatable work benefits from product-spec-like requirements; flexible input asks for missing details but can keep gathering context if the user does not answer.
Sun calls Astra “our most aligned model,” citing greater resistance to prompt injection and unintended actions. That matters for procurement, vendor interactions and customer communications. Yet enterprise reassurance also involves showing administrators internal and third-party evaluations and giving them explicit ways to exclude sensitive applications or data sources.
The deterministic layer is a configurable sandbox: administrators can permit particular applications and websites while excluding others. Separate approval settings can require a person’s intervention for downloads, uploads or entry to a new website. Sun describes these as “dials” an enterprise can adjust, allowing different workflows to carry different levels of autonomy.
Sun describes a system called Guardian—or “auto-approve” in the app, he thinks—that uses a model to assess each action for safety and consistency with the original intent. Consumers may want minimal interruptions; enterprises can choose which actions proceed. Weinstein adds visible cursor movement and permissions for new applications or websites, depending on mode, as ways to build trust before users become comfortable delegating more.
3. Universal access connects fragmented systems of record
Weinstein’s architectural point is universality: “computer use works with everything,” without waiting for a developer to create a connector or MCP server. Sun calls it a “universal kind of API” that increasingly appears inside complex tasks. His desired “shared intelligence layer” would let applications benefit from context gathered elsewhere, avoiding a fragmented world in which every website has an agent that cannot talk to yours.
For the “boring is opportunity” category, Sun’s sharpest example is procurement. Teams fill a spreadsheet with 1,000 rows of lab-equipment requests; operations transfers those requests into internal purchasing software; procurement then uses its output to contact vendors. He sees computer use automating the entire sequence across the spreadsheet, internal system and external interactions.
Sun also highlights month-end reconciliation across expense and other ledgers whose software lacks APIs: agents can reconcile records, flag anything interesting and investigate. His second category is more ambitious creation—circuit design, architecture, hardware engineering and game design—supported by stronger visual capabilities. He thinks usage there could expand as enterprises understand the models beyond the first opportunities to automate tedious work.
4. Proactivity helps when tasks change shape
Sun’s Dots example starts with notification fatigue: his old Codex automation checked Slack every hour or 30 minutes, but inconsistent relevance taught him to ignore notifications. A Dot can instead follow a judgment-based goal—flag messages from his reporting chain or critical product feedback—when the task is “not the same shape every single time.” Each Dot gets its own cloud computer, letting it act on what it finds.
For a relevant bug report, Sun describes the agent using a browser to enter Sentry, retrieve logs, analyze them and return feedback. That replaces his manual loop of deciding which report matters, opening Sentry, downloading logs and feeding them into Codex. He says Dots can control both local and cloud computers, supporting this kind of proactive, long-running workflow.
Weinstein’s consumer examples include DMV forms, unsubscribing from emails, and work people previously lacked time to undertake. His own examples include finance and meal-planning spreadsheets in Apple Numbers, which he calls an “obscure spreadsheet program” that the model nevertheless handles well. He also mentions a personal trainer making customized materials for clients. These are tasks that can be time-prohibitive in personal life but become easier with computer use.
5. Legacy systems and crafted interfaces can coexist
Sun asks whether agents preserve clunky incumbents or enable new AI-native vendors. Weinstein expects both: the cost and complexity of creating software are much lower, but production systems written 10, 20, 30 or 40 years ago will not disappear overnight. Computer use “may put more pressure” on vendors to work better with agents because their software can already be operated through its existing interface.
On the open web, Sun identifies a business-model conflict: he thinks agent ad impressions usually do not count toward revenue. For willing sites, OpenAI supports WebMCP in its in-app browser. He imagines people retaining enjoyable hotel or shopping browsing, while agents handle tedious work; the person can finish the transaction, enter their credit card, and then hand work back to the agent. WebMCP lets the person and agent use the site’s interface together.
Sun’s defense of craftsmanship is a tactical trainer he built for his wife for a game of connecting five pebbles that he thinks is called Omok. He tailored the learning experience to her preferences; an agent could summarize mistakes afterward, but playing remains her activity. “I don’t think UI is going away.” Weinstein extends that possibility to software designed for “100 people or one person.”
Serafini’s enterprise counterpart is an interface for setting goals and inspecting logic. A user should be able to understand why a payment happened and specify timing or a discount at 30 days without executing the whole workflow. “Why should we settle for shitty software?” Weinstein imagines purpose-built interfaces with agents underneath, sometimes operating legacy applications to deliver the experience.
6. Startup opportunity remains in domain knowledge and deployment
Sun remains bullish on startups despite OpenAI’s scale: even with thousands of employees, its teams feel stretched. Founders who understand an industry’s regulation, go-to-market motion and customer service still have room to build. His formulation is that “speed and pain are kind of the moat,” with slower AI diffusion across many verticals leaving substantial work available.
His restaurant example joins several systems: photos of back-of-house storage help estimate what will be missing based on projected traffic for the next day; computer use then orders groceries through Shopee or another procurement application. The workflow also adjusts employee schedules in an ERP, taking account of historical traffic, such as a holiday period. Sun’s point is that these customers need help understanding the tools, connecting them, deploying them and servicing them.
Weinstein’s practical advice is to decide whether the product assists a person at their desktop or operates software on a cloud machine through the API. Human intervention should follow that choice. Founders must also decide when to report success or failure and how to summarize activity: a stream of “this happened” notifications will eventually be ignored.
7. Cloud distribution needs a more coherent experience
Weinstein describes six to nine months of change: capabilities improved earlier in the year, followed by rapid product uptake in the last couple of months, although only part of OpenAI’s user base uses computer use. Sun’s candid lesson comes from Atlas, which he joined in January: “RIP Atlas.” Its agentic features were not a major draw because many people treated it as an ordinary browser.
Sun says incorporating browser capabilities into existing products with established agentic usage produced “YC-level growth week over week,” without giving numbers. He sees cloud access as the next inflection: giving ChatGPT users, including people using their phones for daily life, a powerful computer they do not need to understand how to operate is his route toward a billion users.
For the next quarter, Sun emphasizes coherence. The Chrome extension can use Chrome’s password manager and other Chrome features, while the desktop app’s in-app browser has different cookies and passwords, and the cloud computer has different things. Consumers encounter several browsers that do not communicate as expected. Computer use is conceptually simple—“it’s an AI using your computer”—but those implementation details make the experience harder to understand.
Weinstein wants to add real-time, multimodal collaboration to a product currently strong at background work. He also identifies a behavioral hurdle: people must remember to delegate and change established habits. Dots’ proactivity suggests a way forward, with the agent identifying useful work itself. His ambition is for it to “adapt to you instead of you adapting to it,” while people continue setting and steering goals.
Full transcript
We’ve gotten computer use to a place where it’s faster than the average human, and I think we’re getting really close to where it’s superhuman.
What sorts of tasks do you think are ready to hand over to the agent, versus the ones where it still requires some checking and correcting?
Computer-using agents, since Astra, are now capable of doing anything a person can do. We think a lot about what new consumer experiences we can enable with a faster, better model. On the enterprise side, that’s naturally why safety becomes very interesting. How do we give people the ability to control the amount of human-in-the-loop involvement, and ensure auditability?
It’s unlike anything else we’ve built, where the developer has to do some work to adopt it; someone has to build a connector or MCP server. Computer use works with everything.
Now we have capable models. They’re getting cheap enough to run at scale. What should companies think of automating?
I think of it in 2 big categories. The first category is—
I’m Seema Amble. I’m joined by my colleague, Fabrizio Serafini. In August, we wrote about computer-using agents and how they’re starting to do the real work inside companies. Today, I’m joined by Ari Weinstein and James Sun from OpenAI, who are going to help unpack what’s going on with computer-using agents from their perspective and share some thoughts on the future. Ari, James, thanks for joining us.
Of course. Happy to be here.
Yeah, thanks for having us.
Ari, kick it off and tell us a little bit about what in Astra has made computer-using agents move forward, and what’s really unlocked.
Yeah. Astra’s an amazing model. It’s good at so many things. Our computer-use models were already so good before Astra. Sol itself is amazing at computer use. Astra helps with a few things around the edges, in my anecdotal experience. One is that it’s a little bit faster on average because it uses fewer thinking tokens to do the same task. It’s better at certain kinds of creative work. For example, in our launch video, you were talking to it while it was drawing things in Microsoft Paint. It’s way better at that sort of thing, which I think probably transfers to other kinds of creative tasks, too.
The last thing I find is that it’s just more reliable around the edges. I feel like I trust it a little more. There’s a little difference that’s hard to perceive but really important between something being 99% reliable and 97% reliable. I feel like it’s just a little higher in a way that feels pretty concrete.
Yeah, and adding on to some of those things, Astra’s also our most aligned model.
Yeah.
Right? And talking about trust a little bit—
Yeah.
If you look at the system card for it, it’s much better at being resistant to prompt injection. It’s much better at not taking any kind of unintended actions. When you think about it for enterprise work, especially for companies using it for sensitive things, whether it’s procurement, talking to vendors, or talking to customers, you’re going to feel a lot safer using it. That’s really exciting.
On the creative side, internally we had Astra play video games against 5.6, and it wins every time. There are some cool tweets where Astra was able to beat Fallout in about 30 hours or something like that. It’s also just really amazing at playing video games. I know that probably translates to something, but it’s very exciting.
Maybe just to double-click on that point: consumer versus enterprise use cases of computer agents have very different implications around all the scaffolding around computer-using agents. How do you think about where things are more exciting between consumer and enterprise, if you were to pick?
Our goal at OpenAI is really about building products that are useful for everyone. Ari is definitely more of a consumer guy here, and I spend a little bit more time on the enterprise side. I think the consumer is always the place where we try to push the frontier when it’s possible. With Astra, like Ari mentioned, it’s much faster compared to the previous models. We think a lot about what new consumer experiences we can enable with a faster, better model.
On the enterprise side, that’s naturally why safety becomes very interesting. When we talk to large companies, the use cases for computer use are actually fairly obvious. Inside enterprises, there’s lots of legacy software that doesn’t have any MCP tools. It’s software that’s been used for 15 or 20 years, and a lot of times it’s actually kind of hard for a normal person to use.
Right.
There’s lots of UI from the last decade, and you don’t even know how to operate some of the key things. We’ve seen companies build skills that teach the model to use them better. Maybe the way to think about it is that enterprise is where the capabilities have been there for a little bit longer, and companies are catching up but trying to understand how to deal with it. For consumers, we’re trying to figure out the next types of use cases that may not be possible today with software. Maybe you can share a little more.
The consumer use cases are so fun for me because it’s about how you can put your agent to work doing the things in your daily life that are so annoying, save you tons of time, and give you more time to focus on the things you’re excited about. It’s been really fun to see Dot, where each Dot gets its own computer and you can ask it to do things for you in the cloud. It gets a full web browser. It can buy things for you, book haircut appointments, and do anything else you want it to do. That’s been really fun, and I think there’s a lot more on the horizon that’s going to be really fun.
That was going to be my question about Dots. James, I think you probably use a Dot to walk through product feedback. It’d be cool if you could share more about it. I’m curious to know how you’re using Dots and how computer use plays into what you can do with them.
I think Dots is really cool for a few reasons. One is that it’s really proactive. I’ve been surprised a lot of times, both in my personal life and at work, when Dots flags something that completely slipped through the cracks. Just the other day, internally, our email system didn’t show external invites on our calendars, so I ended up missing some meetings. Dots is now catching these things for me. It’ll say, “There’s this invite in your inbox. It’s not on your calendar. Do you want me to accept it in Gmail so it shows up?” That’s been a lifesaver. It’s a small thing, but it matters a lot.
I also have it proactively looking for tasks where the shape of the task isn’t super concrete every single time. For example, we use Slack internally, and there are a ton of messages going on at all times. I used to have a Slack job using automations in Codex that would go and try to pull things every hour or every 30 minutes. I ended up tuning out a lot of the notifications because sometimes they were useful and sometimes they weren’t.
Right.
With Dots, I can give it a high-level goal: “If anybody in my reporting chain says anything, flag it to me. If anybody is giving critical feedback about certain products, use your judgment and go for it.” Because it has a cloud computer, I’ve had this pretty cool workflow where, when someone reports something, they usually have a Sentry log. I can have it use the cloud computer to drive my local browser. The Dot can control both your local computer and the cloud computer, log into Sentry, pull the logs, analyze them, and give me feedback.
It’s saving that entire loop of figuring out which feedback is important, looking at it, going to Sentry, downloading the logs, feeding them into Codex, and asking for thoughts. All that can now just be automated. Dots is great for these proactive and long-running tasks where they’re not the same shape every single time.
We talked a little bit about reliability, speed, token usage, and efficiency.
How do you guys think about balancing all of these things as you keep making improvements on what you're releasing?
I think we're really excited about computer use at OpenAI, and we've invested in it a lot. We've gotten to a place where I think we have the fastest, cheapest, and most reliable computer use product. There's actually so much further we can take that. We spend a lot of time looking at speed.
Speed is really interesting because, on one hand, the power of computer use is that it can do tasks in the background, so in some cases, it doesn't matter how long it takes because you can be going and doing other things while it's accomplishing the task for you. But sometimes you want to watch it to make sure it does it correctly. You need to wait for it because it needs to complete this before something else can happen.
Speed is so important because you don't want to get the feeling of, “Oh, I could've just done that myself, and it would've been so much easier.” We've gotten computer use 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's superhuman, meaning it's faster than an expert computer user, like probably one of us at this table. I think that's going to be a sea change when it happens. It's going to change the way we all interact with computers, because all of a sudden, anything that you could do manually, it could do faster.
Unless you really want to do it yourself for some reason—which I think there may be many reasons people would want to do that—in the default case, you probably just want the agent to do it for you. I think that's going to let us build all kinds of cool user experiences. Speed is something we think about a lot. Obviously, we have a strategy of offering multiple different models at multiple different qualities and price points, so for different use cases, you may want to use a different model. At the same time, our less expensive models are getting much better at computer use over time, so I think there will be more variation over time in which models people want to use for which tasks.
I think that's what I was saying. Another thought is: how do we give customers control over safety, reliability, autonomy versus speed?
Mm-hmm.
When the model is slow, you can watch it do stuff, but then you're limited to the tasks you can do because you can't just sit here watching it. It's actually less effective, because it's slower than you. But once you 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 oversight there is—
Right.
—and then the auditability?
Mm-hmm.
As part of Astra, we launched a set of enterprise controls that let businesses choose how much human-in-the-loop oversight there is. Do they want the agent to automatically approve everything? We have this really awesome system called Guardian—or auto-approve in the app, I think—which is a really smart model looking at every single action the computer use agent is taking to decide if it's safe and aligned with the original intent.
A consumer might be like, “I just want it to do everything, and I don't want to be asked for approval.” An enterprise can decide which types of actions actually go through. I think the control part becomes more and more important as our computer use agents also become superhuman at doing things.
Totally. As you try to roll out more enterprise use cases, employers are excited about the efficiency, but they're also wary of—
Yeah.
—what's sort of being unleashed. Maybe on the other side, have you guys seen the response from application software? Now there are computer-using agents accessing my legacy workflow software. Do you guys think about that at all? Are you more focused on just the agent being used by the consumer or enterprise user?
Mm-hmm.
Hmm.
Yeah, I think this problem becomes really interesting on the open web, actually. When it's within software that's approved and used by the company, it's more straightforward—
Right.
—because it's like, “Oh yeah, 100%, we want people to automate and do work in here that they don't have to do tediously themselves.” But when it comes to the internet, I think you have a few different camps of folks.
One is: our business model is not in line with having agents use our product. Think about advertisers and publishers whose business model is entirely based on advertising. If the agent's using the website for them, where's the ad revenue coming from?
Right.
I think in most places, an agent's ad impression does not actually count toward real revenue. That's an interesting question that we're always working through and working with our partnerships team on.
There's another question around companies that really want to embrace having computer use agents use their sites. A lot of them are adopting standards that let humans and agents use them together. WebMCP is a standard that we're really excited about. We added support for it in the in-app browser, so you can imagine scenarios where the agent elevates the experience of using the site but doesn't entirely replace it. Browsing the internet is still very fun, in my opinion.
Yeah.
Right? I still like shopping. I still like looking at all the hotels and flights and places. You can imagine places where the agent does all the tedious work, but carves out a user experience where you're doing the browsing and saying, “Okay, now I want to finish the transaction,” and then putting in my credit card and doing all this stuff. Then you can hand it back to the agent. WebMCP is an interesting way for your interface with the website to still be the site itself versus an API, while the person is using it together with the agent. Those are just some thoughts.
Shifting gears, I want to double-click a little bit on what it means to get the end-to-end job done. I'd love to hear a little bit from you guys about what works well, how you define the right narrow, repeatable task, and what sorts of tasks you think are ready to hand over to the agent versus the ones that still require some checking and correcting on the user side.
Generally, I've found the model to be pretty capable of doing anything that you can do on a computer now. One exercise that I try to force myself to do is, whenever I have a task, the first thing I think about is: can I just have the agent do it? There are obviously some sensitivities around sending a message to someone; I want to feel like it's coming from me.
But in general, the way I think about it is that computer-using agents, since Astra, are capable of doing anything a person can do. Before, I think there were questions about some squishiness around specificity and requirements. I've generally found that the model is getting pretty good at asking you for more details whenever you're asking it to do something and there's some vagueness. With Astra, I think we launched flexible input, right?
Yeah.
It will ask you, “Hey, I want some details,” and if you don't answer it in time, it'll just keep going and then pull the context as necessary. For repeatable workflows, it's nice to be able to define the requirements; it's at least similar to writing a product spec. In general, unless there's something incredibly sensitive that you want to be in the loop for, I just reach for the computer use model for everything I do now. I'm curious if it's the same for you.
Yeah, it is. I'm trying to build that impulse to—
Yeah.
—to reach for it.
Yes.
Because it's so good, and sometimes it's so easy to get locked into—
Yeah.
—the old way of doing things. The really cool thing about computer use as a technology is its universality. It works with every application. It's unlike anything else we've built, where the developer has to do some work to adopt it. Someone has to build a connector or MCP server. Computer use works with everything, and it works by using the application the same way a human uses it, which I think is something worth understanding if you haven't spent much time with this technology.
Computer use uses computers the same way a human does. It gives the agent all the capabilities you have, which I think is so powerful.
Yeah, and adding on to what Ari said, part of it is that over time, I’ve stopped thinking about whether I’m using a computer-use task or a non-computer-use task. They’re frequently connected. The way we think about computer use is that it’s not a thing you pick up and just say, “I want to do something with it.” It’s really the universal kind of API the model has to do whatever it takes.
If you have an ambitious task and you want the agents to do it, the more complex it becomes, the more likely it is to touch computer use at some point along the way.
Mm-hmm.
The only time I don’t reach for computer use is when the task is so simple that it’s easier for me to just click 2 buttons and do it. But as computer use gets faster, the question becomes: maybe I don’t even need to do that. I just drive it with voice, and it uses the computer with me. Those are some interesting thoughts, too.
Do you think there’s anything you have to share with an enterprise user to get them comfortable with that? Obviously, you guys are totally built on computer use, but I think there’s sometimes hesitance from enterprise users. They’ll say, “Okay, the human would click through it and probably do it worse,” but there’s that feeling of accountability. If that person is trained, and all the classic rebuttals—I’m curious what you have to say to convince people.
For sure. We’ve had a lot of conversations with enterprises over the last little bit, and I think the main thing is that we think about it as 2 layers. The first layer is: how do we create deterministic controls? I think that’s really what gives enterprises a feeling of safety. Even though our model is the most aligned ever, and I think it’s better than most people at protecting against prompt injections and unintended things because the intelligence is so smart, there’s still a feeling of safety when enterprise admins can say that certain tasks or data sources are too sensitive to be used.
We launched 2 things. One is the ability to deterministically create a sandbox around your agent, so you can say it can use certain apps and websites, but not others. The second is a human-in-the-loop component: what types of actions require a human to intervene? We might say that any time there’s a download or upload, a person has to approve it. Any time it goes to a new website, a person has to approve it. Those are all dials an enterprise can adjust.
The other layer is showing enterprise admins the evaluations we’ve run internally to demonstrate how aligned these models are, compared with both internal and third-party evaluations.
I think there are also some baseline things we do across the board to help people build trust with these kinds of products. We know it’s so new, such a magical and sometimes alien experience for people. We’ve put a lot of work into the product design, so you can always see the cursor if you want to watch it perform a task. It animates in a fun way, so it’s not boring to watch.
It’s customizable for enterprises, but also, for a regular user, we ask permission before using a new application or website, depending on which mode you’re in. I think that helps people not feel like it’s going to do something runaway on their computer. You want to have this trust-building moment where you figure out what it feels like and what you want it to do for you. Then you feel more comfortable over time.
The Codex cursor is so cute. I love it.
Yeah.
It’s such a small detail, but watching it use the computer gives you confidence, and it makes the experience nice. There’s a user-experience point that’s worth investing in. I’m curious, in a broader sense: now we have capable models, and they’re getting cheap enough to run at scale. What should companies think about automating? Where do you see the most promising use cases within enterprises? Are there workflows around ERP, or around how humans and agents can work together to get tasks done? If you had to pick, where are you most excited to see your models perform in the world?
1. Automating Enterprise Workflows
Yeah. From my perspective—and I’d love to hear Ari’s—I think of it as 2 big categories. The first is automating tedious workflows. That’s the more boring one, but I’ll give you some boring examples.
But boring is opportunity, as you said.
Boring is opportunity, exactly. ERPs are a really big one, along with all the acronyms you can think of: ERPs, CRMs—
All the systems of record.
Yeah, every single system of record. We’ve seen this with procurement software. Say you’re ordering a bunch of lab equipment from vendors every year. Instead of having a giant spreadsheet with 1,000 rows that the operations team has to process, computers can automate the whole thing.
Normally, the teams fill out a giant spreadsheet. Then an operations team takes it and enters it into internal procurement software. That gets finalized as a bunch of line items, and then the procurement team takes the software output and reaches out to vendors. Procurement is a big one we see.
Finance is another interesting category. There are a lot of different ledgers within a company, for expenses and other things, and there’s a lot of reconciliation at the end of the month. A lot of financial software doesn’t have APIs to quickly pull things. The finance team can use computer use to look through all these ledgers, reconcile them, and pull out anything interesting to flag for a person or investigate. That’s another boring but useful workflow.
A lot of sales teams use different software for transcripts, data, and other things. Computer use can do prospecting, put the information into a CRM or outreach software, and handle the rest. All the systems of record are the boring part.
Market research is another one we see a lot. In my free time, I’ve been building video annotation software for badminton. I play a lot of badminton, and I like to film myself playing. I wondered how to make the app really useful. There’s software out there for pickleball, squash, and tennis, so I had computer use log in, create accounts on those platforms, try them, take screenshots, give me feedback, and then help me build that back into the app.
Those are all tedious workflows that could be done. Then there’s more interesting creative work that enables more ambitious projects. We talked about how good the model is now at visual design. Circuit design, architecture, hardware engineering, and game design are areas where we’re really excited about what our models can do. Enterprises are thinking about the boring opportunities right now, but as people understand the capabilities of the model more, I think we’ll see more usage in those areas next. Ari, do you have any thoughts?
Yeah, I think it’s similar on the consumer side. There are a lot of things where it’s automating drudgery.
Or just doing things you don’t want to do. It’s filling out DMV forms for you.
Mm.
It’s helping you unsubscribe from a bunch of emails in your inbox. I’m hearing from a lot of people about doing things they just wouldn’t have had time for or wouldn’t have thought to do otherwise. For me personally, I’m maintaining spreadsheets about my finances and spreadsheets about meal planning.
Yeah.
I have a friend who’s a personal trainer who’s making customized things for his clients. Some of these things would be so time-prohibitive to do in your personal life, but this makes it so easy. I use Apple Numbers, which is an obscure spreadsheet program, but the model still knocks it out of the park.
I don’t think I’ve ever heard someone use Apple Numbers.
It’s surprisingly great spreadsheet software. I think there’s a third category of things that I’m the most excited about, which is just starting and, on the horizon, really about having a different way to use your entire computer or interface with software. I think you saw in our Astra ad really cool examples of people just talking to their computer with their voice. It’s such a natural way to use your computer, and I think over time we’re going to see the form factor of software itself change due to this technology. I’m really excited about that.
2. Software Still Needs a UI
On that point, if you’re an enterprise user, you’re used to using whatever clunky ERP was built 30 years ago. There’s a world in which you now unleash a set of agents that just operate that software, and you don’t really care. There’s another world in which a new AI-native player is using a computer-using agent. Do you think the world is going to go more in the direction of keeping the legacy folks and just having this layer on top? Or do you think computer-using agents can absolutely be part of new software players? This is something I think a lot about, so I’m curious to get your guys’ thoughts.
It’s such an interesting moment where simultaneously the cost and complexity of creating new software are so much lower, so you can imagine software changing so much. But also, there’s so much legacy software in the world that computer use is so well-suited to interact with. I expect both of those things to continue for a long time. If you look at the software that’s in production today, a lot of it was written 10, 20, 30, or 40 years ago, and I don’t think that stuff’s going to go away overnight.
I do think that, in some way, the existence of computer use may put more pressure on companies to make software work better with agents, because all the software is going to work with agents by default. I think WebMCP is a great example of this, and I think we’ll see more stuff like that. I sort of expect both to be super important, but I do think it’ll shift over time. What do you think, James?
I agree with you. There’s a lot of art and craft, and you understand this because Ari builds a beautiful product with his team. Cursor was all him. I think there’s a lot of art and craft in building great products that people can hold in their hand. For me, if the internet becomes this faceless thing that’s just agents talking to each other and people aren’t using anything anymore, I would be very sad.
I don’t think it will be like that, because there’s something about craftsmanship and art in software that just feels nice to use, because the person who designed it really understands the experience the customer wants to have. My wife recently got into playing this board game. I forgot the name of it. It’s where you connect five pebbles to win. I think the name of it is Omok. Anyway, she’s trying to get good at it because she’s playing online with friends and getting her ass kicked. So I was like, “All right, I’m going to build you a tactical trainer for going through different gambits so you can get better at it.” I created a first version. I knew she likes to do it in a certain way; she likes these types of ways to learn more than others. So I was able to craft the software to feel really good for her.
Mm.
These kinds of personalized software—I know today it maybe still feels like only tech people will do it. But if you think about this a few years out, anybody can build anything, and they can build this really crafted software in a very short amount of time for niche use cases that are so well-tailored for people. I think that’s a beautiful world. At the end of her training, an agent could summarize the lessons and mistakes she made and give them to her in a text format. But when she’s actually playing with the AI and trying to get better, that’s something the person has to do. So I’m bullish that I don’t think UI is going away.
I think, a little tangentially to your original question, that the pressure of cheap software will make it so software has to be better. But I don’t think it’ll just be that everything will have an MCP and an API, and computer use will never be useful.
Totally.
I love this point so much. Humans are so hardwired for direct manipulation. You really want to be able to move and do things with your own hands.
Yeah.
I think it’s really exciting to think about how software in the past was so one-size-fits-all—
Mm-hmm.
Yeah.
—and had to be. A lot of the software we interact with every day is designed for millions, tens of millions, hundreds of millions, or billions of people to use. I think we will enter an era where any piece of software could be designed for 100 people or one person, and that kind of software will be so much more personal and unique.
My mental model is that there are probably 2 buckets. On the consumer side, you’ve got either something you get joy out of—you enjoy using it or you learn from it. It’s not the utility of it; it’s the joy of interacting or learning or whatever it is, and you want to spend time. On the B2B side, you may want your agent to complete a bunch of the tedious work, but there’s also this concept of the audit log.
That could be personalized in how you want to interact, but the product really is me as the human user setting the goal and being able to understand the logic, but not necessarily needing to run through the entire chain of command of how it happened. Any B2B user still needs to figure out, “Okay, why did that payment get made?” and go back in and get a summary of what happened, and be able to say, “Okay, pay this person at this time, at 30 days, this much discount should be applied,” et cetera. They want the agent to take care of the execution of it, but they want some level of control still to be put in there. I think that does require a UI of some sort.
Yeah.
Yeah.
If we’re moving more toward believing that AI will do more of this, that part has to be enjoyable to use too, right?
Yeah.
Right. Yeah.
Why not? Why should we settle for shitty software?
Yeah.
Yeah.
You know, it’s—
I do think it’ll be really interesting to see, and I’m optimistic that we will see new kinds of software where it has traditional user interfaces but is also using agents under the hood. Maybe sometimes those agents are themselves using computer use to automate legacy software, giving you a much nicer, more purpose-built experience. I’m really excited for that class of thing.
Yeah.
Yeah. It’s kind of like the one we like to use internally: the shared intelligence layer. If you have your agent using legacy software somewhere else and doing things at other places, you…
Like, if you’re using an application, the application wants access to that context so it can deliver—
Right.
—a better experience. If every website has its own agent that doesn’t talk to your agent, it’s a very fragmented world. Figuring out how that plays together—where the agent can plug in and use your other software—I think is going to be a really interesting thing to play out over the next few months.
I love that you guys talk about this, because I think the importance of human-AI interaction is so underappreciated, especially if you want a product to be successful. You have to be thoughtful about how an agent can fit into a team, because humans are used to collaborating. I’m curious: if you apply that to a new startup—let’s say you’re starting a company tomorrow and want to use Computer Use to solve a workflow within an enterprise—how would you think about it? What are the things that jump to mind? You want to know when it fails; you want it to communicate with people; or maybe you want it to be a black box. If you had to talk to founders right now, what advice would you give them as they try to use your models and capabilities to solve problems for enterprises?
In my mind, it depends so much on the modality. Are you a small team that wants individuals on the team to be able to use some piece of software more efficiently? Or are you building something that itself wants to interact with existing software, in which case maybe you have a machine in the cloud somewhere doing Computer Use in the background? That will really affect the modality of the human-agent collaboration.
In the former case, a lot of our products serve this use case. You sit down at ChatGPT, and there’s a lot of great collaboration you can have there. In the other case, maybe you’re building something with the API and doing your own Computer Use implementation. Then it’s interesting to figure out when you want the human in the loop. Is the agent telling you when something’s wrong? Is it telling you when something went right? Is something summarizing it so you don’t get bombarded with notifications saying, “Hey, this happened, this happened, this happened,” which you’re just going to ignore? What do you think, James?
Both Ari and I were startup founders before we began OpenAI, so I’m still very bullish on startups. The question comes down to: Is there still a moat? Is there something that startups can still tackle while the labs are doing things? In general, I think there are 2 things. One is that there are so many things to build. Even though we’re quite big now, with thousands of people, we still feel very stretched thin internally.
Oh, yeah.
The teams aren’t very big for all the problems. If you have a problem that you deeply understand, industry expertise, and you understand regulation, the go-to-market motion, and how to service those customers, it’s still a very green field out there. The tools you have through our APIs, like Ari mentioned with Computer Use, will still let you accelerate. I think speed and pain are kind of the moat that exists today.
The other thing I think about a lot is that there are still a lot of vertical industries where the diffusion of AI is much slower. I was talking to a company the other day that uses Computer Use to help restaurant groups take pictures of their back-of-house storage. Vision looks at what’s going to be missing based on a projection for tomorrow’s traffic, then it uses Computer Use to go to Shopee or some kind of procurement software and order groceries for the next day. Then it goes into the ERP to adjust the timetable for employees, who might need more people to help because at this time last year, there was a lot more traffic coming in for a holiday or something like that.
Those restaurant groups probably aren’t tech companies that are super familiar with all the tools available. Living in San Francisco, we hear a lot of thinking like, “Wow, the labs can do everything. Codex—and why doesn’t everybody just install Codex and have it do stuff?” But for the average American, the average small or midsized business owner, or the average enterprise, understanding what the tools are capable of is one thing; connecting tools in a way that makes sense is another; deploying them is a third; and servicing them is a fourth.
There’s still a lot of opportunity, in my opinion, when complex workflows require a lot of disparate software. Computer Use is a great way to join them and complete the entire end-to-end workflow. Building it in a format and UI that feels good to use for those customers still feels like a really big opportunity that remains open.
Right. Tell us about what’s next. What are you excited about on the frontier that you can share around Computer Use and agents?
Ari?
I’ve sort of gestured toward it already, but we’re really excited about speed. We’re really excited about making Computer Use superhuman, and we’re really excited about things that blend together different kinds of interfaces and bring together multimodal inputs and outputs, so you can have more fluid experiences with your computer.
Right now, a lot of our Computer Use product is very background-focused. It’s very much like, “Let me give the agent a task so it can do that while I do something else,” which is actually a tremendous strength of the product. But I think that in the future we’ll also see more real-time collaboration between you and the agent. I’m really excited about this area. Otherwise, I’m excited about things being better, cheaper, more available, and easier to use.
More. Yeah, exactly.
Yeah.
Being able to actually access everything.
Yeah.
With Codex, we’re giving folks these computers in the cloud. ChatGPT Work also has a browser in the cloud. For me, one thing I’m excited about is making the experience feel a little more coherent.
Mm-hmm.
Today, you can install the Chrome extension, which uses your password manager and everything in Chrome. Then we have an in-app browser in the desktop app, which has a different set of cookies and different passwords. Your computer in the cloud also has different things. For the average consumer, it’s kind of confusing.
It’s like, “Why do I have these 4 different browsers that don’t talk to each other? If I open a tab in one, does it show up in the other?” A lot of what I’m excited about is how we make Computer Use go from a product used by enterprises and really hardcore tip-of-the-spear users to something that’s understandable for people. It’s kind of an oxymoron, because I think Computer Use is actually the easiest-to-understand AI product—
Mm-hmm.
—because it’s an AI using your computer.
Yeah.
But because of the complexities of the different states you have to manage and the different places you’re using it, the implementation details sometimes seep through, and then it becomes kind of hard to grok. Being able to make that consumer-grade, easier to understand, and more coherent is something I’m really excited about for the next quarter.
Yeah, I think it comes back to the point that there is still a human setting off the agent, right?
Yeah.
And the better the UI and controls are, and the easier and more enjoyable it is to use, the more they’re going to get out of it. That’s still the principle.
Absolutely. On that point, one other thing I want to gesture toward is that I think another challenge with the product today is that you have to put some thought into how you’re going to use it—
Mm-hmm.
Right.
And you have to remember to use it, and you have to convince yourself to do things that are different from how you're used to doing them. I do think Dots, like James was talking about earlier, have this really powerful proactivity where they can come to you, identify situations, and help you solve problems. I think we'll also see a lot more proactivity. I'm really excited about a world where the agent can adapt to you instead of you adapting to it.
Mm.
Mm.
I think we're going to see a lot of that, and I think it will make it a lot easier for people to relate to, understand, and incorporate this technology, while still being the ones setting and steering the goals, if that makes sense.
One of the things we haven't talked about yet is the momentum behind Computer Use and how much has changed. Fabrizio and I chatted a little bit about this in our piece. I'd love for you to talk about where things are, how the trajectory has gone so far, and where things are headed.
Mm.
Yeah. Yeah.
At OpenAI, James and I have been working on Computer Use for some time now, and it's been an incredibly exciting past 6 to 9 months, right? We started working on these Computer Use products, and very few people were using them. The models all of a sudden, it felt to me—
Mm-hmm.
—became great at these kinds of tasks—
Yeah.
—earlier this year because of amazing work the research team did.
Totally.
Then we've been working super hard with a bunch of teams internally on how to create good product experiences around this. In just the past couple of months, the product's been growing like crazy, and people are really seeing the value in it. But I think it's still so early, you know?
Yeah.
It's still the case that only some portion of our user base is using Computer Use, and I think we're going to see it grow quite a bit.
Yeah, totally. I joined OpenAI in January, working on the Atlas product, and I love Atlas. RIP Atlas. One of the things we realized was that, in the Atlas product, the agentic experience, I think for a few reasons, wasn't actually a big selling point. Not many people used it, partly because most people treated it like a normal browser.
Over the last few months, because the model capabilities have gotten better, we've incorporated browser and browser use into all of our existing products where there was already agentic usage. We've just seen YC-level growth week over week, and I'm always surprised by how fast it can go. I think this next big inflection point is cloud, right? How do we—
Right.
—how do you give every single ChatGPT user, everybody who's just using their phone for their day-to-day life, a powerful computer that they don't even need to understand how to use, that just makes things work for them? I'm really excited about how that brings it to a billion users.
It feels like computer-using agents have had a ton of momentum behind them, and now it's the moment of, “Okay, how do we get it into more people's hands?”
Yeah.
They trust it.
Totally.
They feel like they can really use it. And to your point earlier about coherent experiences—
Yeah.
—once that gets a lot better, I think consumers will feel more comfortable using it, and it'll feel more natural.
I think that's right. I think we're at the scaling moment of Computer Use, both for consumer and enterprise, at the same time. I'm really excited to see where it goes and what people build with it.
Nice. Us too.
All right.
Well, thank you so much for joining us. This was awesome. We love talking about this topic, and it was really fun to hear it from you guys directly.
Thank you. We had a lot of fun as well. Yeah, this was really fun. Yeah, thanks. Thank you. Thank you, guys.