A billion people every single week use ChatGPT. Like a billion people. No one else is doing that in human-operated software. The response is too slow. I feel like this can be the most empowering technology that is ever created.
I think it would be helpful to start with just the news this week: Denise Dresser, your CRO, leaving the company; you bringing in a new one; Brad Lightcap leaving. There are a lot of questions publicly about OpenAI's leadership bench and what's going on, and you've come in and are now managing a lot of this, and I'd love to hear from you. Are all these departures connected? Are they different? Is there any through line there?
Look, the way that I think about it is that the theme of this year has really been focus. We have line of sight to such incredible technology, but it's not enough just to produce the technology. We need to really figure out how do we have this technology benefit people? How do we bring it into every enterprise in the world in a way that actually uplifts what they're doing, helps them accomplish their goals? So that's really been the singular focus that we've been pursuing. In the pursuit of that, we've been really trying to think about do we have the right structure, do we have the right strategy, do we have all the right things there. My view of OpenAI is that we're a very resilient organization, but a lot of what we've been doing is really trying to build the leadership team that can really help us achieve that goal. My view of everyone who has been here and people who are no longer here is that they've all helped to contribute to what OpenAI is today. Looking forward, I think that we are scaling systematically to every enterprise in the world. We're bringing our models to consumers everywhere, and a lot of my focus has been making sure that we can achieve that.
On the CRO side specifically, what is it that you weren't getting that you needed? Is that the way to think about it?
Look, I would say it differently. A year ago, I think that we really were not in the game at all when it came to enterprise. It just wasn't really our focus. We hadn't really built the muscle we needed. And now, I think we actually are in the game. We kind of have really built a foundation that I think is actually something we can really build on. And now we're systematically scaling. To me, that is like the moment. It's not really about the individuals in some ways; it's really about the overall sort of direction of the field.
How does that change this systematically scaling you're talking about? What does that mean practically for OpenAI?
Well, number one is that I think that we are looking at transforming the entire economy. Like that's the thing that I think is very hard to see from the outside, right? It kind of feels like, "Hey, it's a business, like we're selling SaaS software." That's not what we're doing. We're doing something very different. We are transforming the economy to be a compute-powered one, and I think that what an enterprise even is will change. And so the question is, well, where do you start? How do you make sure that you understand every segment and you're able to sell in a repeatable fashion to all of them? So a lot of the motion that we've been working on—and again, this isn't a new motion in some ways, it's just one that we need to now scale just given where the business is at—is to really make sure that everyone in the field, we're able to funnel back the insights that we're seeing from customers in a very systematic way back to our product teams, our research teams, and vice versa. One thing that we have learned, if you look at, for example, how we produce blog posts internally, we learned years ago that it has to be a collaboration. You cannot have a handoff of, "Oh, here's the comms people who write blog posts and here's the research people who produce models," and you know that they'll just kind of hand it off in some way. It has to be deep collaboration. You need to have the technical expert really hold the pen and feel accountability for the end result being good, and they can be amplified by people who are experts on how do you actually craft a narrative, how do you sort of bring together all these assets, how do you have a good launch?
And so that to me is the core activity that we are pursuing, is to say we can't be fragmented. We can't be working on all the things. We have to be unified and focused. So that's the internal answer for how we're working. In terms of the external, I think that you should see that we will systematically be prioritizing different industries, really focused on the deep understanding of the metrics and exactly where the token spend is growing the most, and really focused on outcomes. Like every business right now is asking, "Where's the ROI?" And that is something that is a big focus of ours. Ours has always been the, "Hey, we have the best price-performant models," but we really have to show people that in the outcomes for them, also really scaling further through channels and through partners. That's something we believe in as well. So I think that a lot of where we're at is we feel very confident about the fundamentals of the motion and the fundamentals of we have a great product, a great model, and then really just rapidly growing both our internal team so that we can reach everyone, but doing that in a way that really has that positive effect that we're looking for.
You use the word accountability, and I'm curious from your perspective, what are you accountable for now at OpenAI because of the leadership changes that have happened? Fidji Simo came in for a time. She had to step back permanently due to her health. You stepped in while she was out and took over a large part of the company. Now it's you and Sam effectively running it together. How is that divided up?
So for the past couple years, I've been accountable for the infrastructure side of what we're doing: data centers, compute, those kinds of things. That part remains. There's a large part of the scope that I was also working on in terms of the software for kernels and sort of how you train these big models, that's something that I've been able to hand off so that I can really focus on the business side. And so I would view me as sort of end-to-end accountable for revenue, for product, marketing, the whole machine for bringing these models from research to value for our customers.
When you look at this machine, and I know you're still putting pieces in place, but when you look at this machine you have in front of you that you're accountable for, how is it doing relative to competition? I mean, I think specifically Anthropic, there's reporting that their ARR is potentially more than double what OpenAI's is. What do you make of that?
Well, I think we started late. Specifically, the way to think about the overall field is that for the past couple years, the chat market was the main growing one, and the reason for this was simple: GPT-4 was the first model where the results were personalized enough that it made sense for you to ask a question, read the answer, it's interesting to you. And so it very naturally lends itself to a chat bot. That is something we're able to scale and deliver value. Now a billion people every single week use ChatGPT. Like a billion people—that is scale, that is real scale. No one else is doing that. But this year something else happened: the agentic moment. This really was December of last year, when both we and competition had a moment where agents were really able to first start doing software engineering in a significant way. It went from saying that you could speed up your work with maybe a 20% productivity increase to 80% of the work could now be done with an AI. And so suddenly the way that software engineers operated, it inverted—that you became much more of a manager of agents. Since then, the growth has been unprecedented, absolutely unprecedented how these agents have been adopted.
One thing that happened was I think that we focused late on coding in particular. Actually, let me sort of give a different answer as well. There's one simple narrative, which is that we focused less on real-world coding, and we came to that game late because we always had looked at things like the coding competition benchmarks and always had the lead there, but less the, "Okay, but how is a developer really going to use it?" Like the messy real-world codebase, the developer interrupting in the middle, what's the personality of the model—these sort of little last-mile paper cuts that actually make a huge difference in adoption. And similar on the GTM side of really thinking about how do you build relationships with customers, how do you build the kind of observability controls that they need, how do you help them understand what's possible within their context?
All of these were less our singular focus until much later than competition. So I think that we started late, but our trajectory has been very steep. But there's actually even a deeper answer that is less obvious from the outside, and that is one of model capability. For a period of about two years, from when we announced GPT-4o in April of 2024 to December of 2025, we were really squeezing the juice out of the GPT-4o model, that we sort of had a bunch of pieces that we were improving, but fundamentally we were on the same underlying base model, the same underlying pretrain. There's a lot of reasons. That is really because to produce and ship a pretrain is something that requires full integration across the company. It's not any one team; it's about the integration across these teams to think about the inference, to think about the post-training, to make sure you've codesigned the model to be the right size. All of these things you have to get exactly right and have a harmonious way of working. Over the course of 2025, a lot of our focus, a lot of my personal focus, was how do we really get back to shipping new pretrains as a core part of our innovation engine.
This year, you see the results. We're on our third pretrain; we're actually about to ship that.
This is Astra. A lot of the gains that you see are because we have made the model so fundamentally better. So the way to think about it is that there's an underlying technological place where we really were lagging. There was a core part of our innovation machine that we weren't able to fully bring to bear. But now it's different. That is why you see our model capabilities taking off so much. If you look at 56 soul when we launched that best model in the market, you could see the revenue really starting to shift. We grew 20% in our top-line revenue in July, and we see that exponential continuing.
Another thing that's changed, as you were saying with the model capability, is the models are becoming incredibly capable at cyber and at hacking. I think the world is still kind of thinking through the implications of the Hugging Face incident. Essentially what happened—correct me if I'm wrong—is you all accidentally hacked another company without knowing it for a while, and it was a model that you all were training and evaluating. There was a Black Hat video where you guys walked through this, and you're going to put more out in the coming days. But walk me through that: when you heard that that had happened, what did you think?
Well, in the words of someone on our security team, it was definitely the most eye-opening security event that any of us were aware of. The way to think about it is that there's a lot of different lessons that we're taking away from it, but I think that it is a moment where we can as a community learn. We can understand what the shape of the future is. To me, the salient points for the purpose of how I'm thinking about it right now is that we had a model and we were evaluating it, and it managed to—it was in a sandbox that was well-configured, and it managed to both break out of the sandbox and then also break into production infrastructure that was intended to be hardened. And so the capability there is quite high. There are some details about exactly how it all worked and the fact that it was kind of not just a singular agent, but really a swarm of agents that were focused on this goal and able to pass messages to each other. The sort of thing that you start to see is that with very capable agents, there are limitations that human attackers have that slow us down that will no longer hold back the attackers, that they can look through lots of information through many different little vulnerabilities and chain them together into something significant.
For me, what I immediately started to realize and think about is, what are the implications for six months from now when this kind of capability isn't just in the lab, isn't just in a few places, but is really widely available? I think that the answer is that every company needs to respond, needs to upgrade its security because attackers are going to fully automate. You can even see with the release of GLM-5.3 that the cyber capabilities of open-weight models available to everyone are going to increase. The important thing is that I actually think in the long run that defenders can win, that AI can actually make it so defenders can decisively win in a way that was never possible before. Because security is always a cat-and-mouse game: attackers uplevel, defenders uplevel. But AI can give defenders an unfair advantage. For example, formal verification of software where you can actually prove that it is secure, mathematically prove security properties. The technology for this exists, but it is intractable for humans to actually do the work to prove those properties.
And we have these models that are incredible at math, right? For example, Astra, our upcoming model—we announced this a few weeks ago—solved 10 math problems, and this was with $2,000 worth of compute, that had been open for decades, mathematicians that spent their whole careers on these problems, and now we have a model that can prove the answers. Those models definitely can prove formal safety properties of software. So that's one place where there's an unfair advantage for defenders. A second thing that we're doing is we're really investing in generating secure code. You start to realize that as a human developer, there are many patterns that people just keep getting wrong. I've been through this many times in my previous jobs; I was very focused on building secure infrastructure and just find all these little things people do wrong. SQL injections people kind of know about, but there's lots of little anti-patterns as well. Even like, for example, I remember it used to be like in Ruby that the regular expressions, there's like certain characters that if you get them wrong, there's just very common attacks where people can put a new line and it doesn't match, so you're able to add things to the payload. All of those problems of just little security anti-patterns that you just keep having to tell your people, "Don't do this, don't do this, don't do this," you can train your AI to never do it.
And so I think that we actually have the ability to build the best security engineers in the world, and fundamentally the defenders can have the advantage, but this is going to require a mindset shift. This is going to require a megaproject. This is something that the whole ecosystem needs to change how software is secured because there's a second thing that we've learned, which is that as these cyber models have gotten more capable, the number of bugs that they are finding in critical software—for example, the Linux kernel—they just keep finding more and more. It really makes you realize that human-written software, the response is too slow. So we need defenders to automate, to move at machine speed, because their adversaries certainly will.
What do you say to people who already have a negative disposition towards AI—maybe there's a data center they don't like in their neighborhood or something—and they hear you say that and go, "You're just selling the answer to the problem you've created," in the sense of like you're building these models and now you're going to sell something to protect me from them? What would you say to that?
First of all, I have a lot of empathy for that reaction. The way that I look at it is I rewind to 10 years ago, 11 years ago, before we started this company. I wasn't in AI. The reason I wanted to be in AI is because I feel like this can be the most empowering technology that was ever created, that AI can be something that helps everyone live the life that they want, build the life that they want, achieve their goals. I just wanted to help steer it in a slightly more positive direction than it would be without me. So I started because I feel like this technology can uplift everyone. When you look at where OpenAI is and what we're doing, we can see a little bit into the future. What we are creating, it's not just us. There's a whole field, there's a whole ecosystem, there are many actors, and that is itself a good thing. We believe that there should be democratization of this technology. It should not be locked up in one place with one entity controlling it, because that's a different kind of risk. But it means that we do need to adapt, we do need to navigate through, and we need to use the information that we have painstakingly extracted from the future to see what is the shape of it. To me, Hugging Face is a sort of proof point, a case study, a window into the future. It's not necessarily a future that any one company is creating; it's the overall economy, the overall set of actors in all these different countries who are building these kinds of models. It shows you what will be possible and how attackers will be able to operate. What it doesn't show you is how do defenders update. But that to me is what we are also trying to now show as well, that we can build these models that you can use to scan yourself.
As an example, I asked Codex—and this is just off-the-shelf Codex, no special access, no special models, I just use my personal Codex—to scan gregbrockman.com, "Tell me what vulnerabilities are there." Gregbrockman.com is very simple; it's just a static site. I was like, "Come on, what's it going to find?" And it came back with like 13 findings. Each one of them, I was like, "You know, I kind of knew about this, but it's so much work to fix it," and like, "Oh yeah, I kind of forgot that I left that open," or, "Here's some domain that was legacy, I'm not using it anymore, I hadn't set the records that prevent people from spoofing your email address," a bunch of little things like that.
That one seems like a big deal.
Well, yeah, you're right, now that you say it. Now, here's the thing: honestly, to fix those kinds of things, you have to set up DMARC to do it right, and there's this whole process you have to do to roll it out. It just takes time, it takes a lot of expertise. So instead I just asked Codex, "Can you fix it?" And so it opened up my Cloudflare control panel and it just used Computer Use. It clicked around, set a bunch of headers, changed a bunch of settings. I used to have it so that I had Cloudflare in front and then AWS on the back end, and it was going over insecure HTTP here. Codex had figured that out. I'm actually not even sure how it saw that because I thought everything was hidden behind Cloudflare. It actually was just like, "You know what? I'm just going to migrate you off of AWS." So it set up Cloudflare Pages, moved over the content. I had an old jQuery version that was insecure; it just got rid of jQuery entirely. It was like 15 minutes to find all these findings, and it was another hour to fix all of them. For me, I just was the manager. I just gave it the intent. I said, "What I want is I want to know where the problems are and I want you to fix them," and it did it. That to me is more the future that we're promised.
I want to hear—and I know this is more in the research world, but I think it's really relevant to everything—you talk about what you guys are doing now in response to Hugging Face. One of your colleagues, Mia, told me the other day, "I wish we had done a lot of the work we're doing before this happened." What is that work that's happening now?
I would view the mode we've been in since the discovery of the Hugging Face incident as full incident response to uplevel every aspect of how we secure, monitor, and align our most capable models. We believe that models that are above a certain capability level require this. It's been a very interesting technical problem to think about, and it's not all us being caught flat-footed because we anticipated this moment for quite some time. If you look back at February publicly, we started talking about trusted access for cyber models. That was the first time we were really thinking about, "Okay, we've always had in our preparedness framework when cyber capabilities get to a certain level, you need to treat them differently." Now we're starting to think about the fact that these are going to be so transformative in the world, we need to think about how access is controlled internally. We started to think about, if we have very capable models that are able to break out of sandboxes, well, we need to make sure our sandbox is as industrial-grade, as foolproof as humanly possible, which really means engineering from the ground up around security properties. We put some of our—actually Mike Dalton, who was one of the people who gave that Black Hat talk, the task that he started to work on was build the most secure possible sandbox. So we've been putting in place a bunch of infrastructure primitives and policy primitives for this moment, but it's also different when it actually hits, right? You never know when you're going to hit that moment. Here we got to see this window in a very concrete way of, "Okay, we're now at a serious capability level relative to the pieces. We must act now."
And so there's been a workflow shift that's been happening very rapidly because a lot of our governance has been focused on deployment of model, but not necessarily the training of model. Now we need to pull the governance back in during the training and evaluation process. We also need good governance there. We need to have standards around checks and balances around how do we know if it's secure enough, how do we know if it's aligned enough. This alignment piece is actually very important because our belief is that if you make models extremely capable, you really want to make sure that as the capability grows, their alignment grows with them, that you want those things to be in lockstep and sort of coordinated fashion. In that way, you have this defense-in-depth of models that are trying to do the right thing, that you have a sandbox that's as secure as possible. Doing this in an iterative fashion so that you're taking incremental jumps so you can kind of understand as you go and make sure that your systems are working, all of that is very aligned with how we've thought about iterative deployment, how we've thought about the creation of this technology from the very beginning. So we are operationalizing that. For me as a manager, thinking about the organizational system for producing this technology, it's also been a very important challenge because this means you now need different people in the room to make decisions. You need a different process for how do you know what the requirements are, how do you execute against those priorities, if there's a change in your technical understanding, how do you propagate that, how do you make sure people are in the right rooms. But again, in some ways it's been a phase shift, but it's been one we've been preparing for. Earlier this year, I spent a lot of effort to really try to bring together research and security to start having—we started doing, for example, these weekly safety and security of agents meetings. Some of these things are about building the connective tissue, having people who wouldn't normally talk to each other because their worlds are a little bit different—they come from different backgrounds, they're focused on different problems—be forced to interact, be forced to think about shared problems. A lot of that shared empathy, that shared understanding, that shared technical context is paying dividends right now.
At the same time, though, the implications of this are that momentum on new models is going to significantly slow. Do you have a sense of the impact?
Well, I would put this differently because I've always taken this lens to how do you think about progress in this field. If you only think about progress as what's your score on a benchmark, what's your overall capability level in this specific task, that's too myopic. We are building an integrated system, and safety and security are a core property of that. So if you have safety and security lagging relative to the capability, yeah, you absolutely need to improve your safety and security. But our goal is to have a system that we are building at full force on safety and security as first-class requirements. The way I would think about this is that we are upleveling every aspect of the machine. We have spent the past year, past two years upleveling how we build our clusters. We now are able to train at massive, massive scale. We're able to get the whole pre-training machine, the RL machine, all of these pieces to work together in concert. We are building the best sandbox. We are adding chain-of-thought monitoring to all of this. To me, it's just saying what is the scope of delivery? What does success look like? What does good look like? So the way I would think about it is that, yeah, if you jettison some requirements, you can move faster in other requirements, but that doesn't make any sense. That never makes any sense. That's not how engineering works. To me, we are pursuing the plan. We've really thought about this many times in the past, and you can see it in many of our writings, from the charter to the preparedness framework and many other things in between, that as models get more capable, you have to make sure that you increase your safety and security requirements in concert, and that's exactly what we're doing.
Practically, yes, I hear all that, but it does seem like what this will result in is the release cadence of models to lengthen.
I'm not certain if that will be true, to be honest. I think that what will happen is it does mean that if you were just maxing a particular metric, you could probably max that metric faster than what we will pursue. But there's a lot of degrees of freedom of, hey, we always have a bunch of wins queued up and we can jettison some of those wins on a particular axis. So if you're pushing on particular capabilities on cyber or on bio or on coding or any of these things, maybe you don't couple together all of the particular things. Or if there's architectural improvements, maybe you delay adding those to the run, but your rate of producing checkpoints doesn't need to change. It's just kind of what those checkpoints represent could be less capable in specific axes, more safety and security intertwined. But I would almost think of checkpoints and our shipping of models as a little bit more about taking a photograph of where we happen to be. It's not like there's a preordained set of models that have to be created; we can codesign. In practice, I think that we actually have a lot of reasons that we prefer to do very incremental releases with these photographs taken with a reasonable frequency.
I guess what I'm getting at is that for a while, the release cadence felt relatively in place for everyone. In the last year especially, it feels like it sped up—new models coming out every six weeks essentially, you and Anthropic putting out new models constantly. What you're describing, this kind of taking stock and introducing more friction into the process, it does feel like that just inevitably has to slow future momentum. I'm curious if you're thinking about that from the perspective of this machine we've been talking about, the business that you're now accountable for, and the impacts on that, or is it too early to know?
I understand that perspective and I think it's a very reasonable one, but my observation of how we operate is that we codesign everything. The way in which the business operates for sure is we will take the most capable model we are ready to release, that we've done the appropriate safety and security controls for, and we will bring that to our customers and help them get the ROI that they're looking for, help them really transform their businesses, get value—simple in a lot of ways. Then the question of what models are ready to release, we have many models that we produce internally, and the choice of how high a bar to set, how low a bar to set in terms of the capability jumps, all of that has to be codesigned: the frequency of it, that we have early access programs, there's programs like Daybreak Red and Blue which let us have different models that go out to different partners. And so I guess the way I would think about it is that there is for sure a shift here in terms of our internal balance of...
effort and where the checks and balances are and how those things operate. I don't think that there is a clear trickle down to just like, okay, the model cadence will go from six weeks to six months, but what those six-weekly releases actually contain will be different.
So maybe the raw kind of frontier doesn't get pulled into production, into the commercial world quite as fast, even though it's still being trained on and worked on.
Well, I would think of this as an overall pacing, right? An overall pacing of that frontier rollback. And another thing to think about is the fact that our ability to produce these models and to get those capability gains is also increasing, right? Our raw capacity is because partly we understand this technology a lot better than we did in the past, and that's actually very beneficial from a safety and alignment perspective too, right? That we really understand a lot of how the technology works.
There's still many pieces that we don't understand, but our ability to engineer it and to really do a good job of it technically, I think, has increased a lot, and that there are certain problems that really used to hold us back that are much diminished.
For example, these big training runs run for many months, and so typically we'd launch them from a branch of code, and so we'd be adding little fixes to that that are specific to that run. At the same time, others continue their mainline development, and every so often there's something on the mainline that we're like, "Oh, we really want that change, we really want that fix." And then we're like, "All right, we got to bring these two branches of development together," and we merge them. And then we launch the run and we find that there's a bug because no one has actually tested all these mainline fixes at the massive scale, because we only have one instance where we can do it. And then we spend the next three to four weeks with all of our best people scouring, trying to figure out where did the bug come from, because you can't possibly read all of the code that's been written, right? No human could possibly do that. Today, you run into that same type of issue, Codex just reads all the code, finds the bug, and says, "Here it is, no problem."
And so there's something about the partnership of this amplification of what the models can do that knocks out whole classes of problems where we know it would take three to four weeks and now it takes a few hours. We've seen this concretely multiple times. And so what I'm getting at is that there are a lot of limitations for how we've had to develop in the past that have made it a lot harder for us to move forward on problems that are important. And now we have the opportunity to move forward on them in a much more efficient fashion. And that to me is like, let's use that extra juice to increase the alignment properties, to increase the safety properties. Let's make sure that we can have models that monitor other models in a trustworthy fashion. Let's really implement some of these ideas that we were talking about back in 2017, right? That we have things like iterated amplification, things like debate where the models can really spend a bunch of compute in order to make sure that what they're doing is actually reasonable. To me, that's the opportunity in front of us.
Do these changes you're talking about impact the release of Astra, the thing that is coming sooner at all? Do you think that's going to be coming later than you guys originally hoped?
Well, it is already later than we hoped in some ways, right? So we've had a model that we think is very good. Lots of people using it internally. It's not out there yet. And so I think that the answer is yes. We have actually been, despite everything I just said, we have very much sacrificed the short-term maximalist if you just wanted to focus on revenue or something like that.
That isn't how we've thought about it at all. No one has even asked that question, because it's just very clear we know what our priorities are. Our priority one is AGI, do right by AGI, and our priority two is transform the economy. So it is on the priority list. Revenue is absolutely a priority for us, but we know the relative prioritization. We know what wins if there's a direct choice that has to be made.
