0:00 [Music] [Applause] Good morning. 32 months ago, we launched Chat GBT and since then it has become the default way that people use AI. In that first week, a million people tried it out and we thought that was pretty incredible. But now about 700 million people use Chat GPT every week and increasingly rely on it to work, to
0:46 learn, for advice, to create, and much more. Today, finally, we're launching GPT5. GPT5 is a major upgrade over GPT4 and a significant step along our path to AGI. Now, today we're going to show you some incredible demos. We'll talk about some performance metrics. But the important point is this. We think you will love using GPT5 much more than any previous AI. It is useful, it is smart, it is fast, and it's intuitive. GPT3 was sort of like talking to a high school student. There were flashes of brilliance, lots of annoyance. Uh, but people started to use it and get some value out of it. With GPT4, maybe it was like talking to a college student. Real intelligence, real utility. But with GPT5, now it's like talking to an
1:32 expert, a legitimate PhD level expert in anything, any area you need on demand that can help you with whatever your goals are. And we are very excited that you'll get to try this. But it's not only asking, though. GPT5 can also do stuff for you. It can write an entire computer program from scratch to help you with whatever you'd like. And we think this idea of software on demand is going to be one of the defining characteristics of the GPT5 era. It can help you plan a party, send invitations, order supplies. It can help you understand your healthcare and make decisions on your journey. It can provide you information to learn about any topic you'd like and much more. This is an incredible superpower on demand that would have been unimaginable at any previous time in history. You get
2:19 access to an entire team of PhD level experts in your pocket helping you with whatever you want to do. And anyone pretty soon will be able to do more than anyone in history could. So today we're going to talk about GPT5. We'll show you some upgrades to chat GBT and we'll talk about the API. GPT5 is great for a lot of things, but we think it's going to be an especially important moment for businesses and developers, and we're very excited to see what they're going to build with this new technology. So, we can't wait for you all to start building with this. We hope you enjoy it as much as we enjoyed building it for you. And to start, I'm going to hand it over to my colleague Mark, our chief research officer, to tell you about GPT5. Thank you.
3:05 Hi, I'm Mark and I'm joined by Max who leads the post training team and Renie from our engineering team. Over the past few years, OpenAI has spearheaded the reasoning paradigm. These are models which pause to think before delivering more intelligent responses. Now reasoning is at the heart of our AGI program and it underlies the technology that we use to ship stuff like chatd agent and deep research. GPD5 aims to bring this breakthrough to everyone. Until now, our users have had to pick between the fast responses of standard GPTs or the slow, more thoughtful responses from our reasoning models. But GPD5, it eliminates this choice. It aims
3:50 to think just the perfect amount to give you the perfect answer. Now, something like this takes a lot of hard work. We've had to do a lot of research to make CHPD5 the most powerful, the most smart, the fastest, the most reliable, and the most robust reasoning model that we've shipped to date. Today, we're going to show a series of demos in coding, in writing, in learning, and in health. But GPD5 isn't limited to these domains. It's very useful in all cases where you require deep reasoning or expert level knowledge in things like math, in physics, even in things like law. And the exciting thing is we're excited to make this available to everyone, even to our free tier. After we show our demos, we're going to be talking about how GPD5 supercharges
4:38 our chat GPD app and our API. We believe that GPD5 is the best coding model on the market today. To start, let's have Max talk a little bit about the benchmarks and how the models stack up. Yeah, thanks Mark. So, as Mark said, we think GBD5 is by far our smartest model ever. So, let's start by talking through some evals. Now, eval aren't everything and they don't tell you everything about a model, but they can highlight its intelligence. And GPG5 performs exceptionally well on a range of academic evals across subjects. It outperforms both our previous models and other models on the market. So picking up first on the theme of coding, GPD5 sets a new high on SWEBench, which is an academic eval that tracks performance on real software engineering tasks. Now
5:23 this again is an eval but we think it will reflect the model's performance in the real world. GPD5 also performs very well on Ader polyglot which measures its ability to implement complex functionality in a variety of different programming languages. Now, beyond coding, GPD5 performs exceptionally well at multimodal reasoning, setting a new high on MMU, actually outperforming both our previous models and most human experts on this task. This is basically a visual reasoning domain where you are asked to from an image figure out what's going on. Uh, GVD5 is also excellent at mathematical reasoning as shown by its performance on Amy 2025. Now this is an exam that American high school 2 students take to qualify for the international international mathematical olympiad and GPD5 performs exceptionally
6:09 well again beating our previous models and other models that are out there. Now moving beyond academic eval towards some real world use cases. We put a lot of work into making GBT the most reliable and accurate model in the world. Language models historically have been plagued by hallucinations, factual errors that make it hard to rely on their outputs for actually important tasks. For GBD5, we made improving factuality, especially on open-ended or complex questions, a priority. We also built a set of new evals to track this, and we're very happy to report that GBD5 is by far our most reliable, most factual model ever. GBD5 also performs exceptionally well on health related questions. Now, health is a big part of how people get value from GPT in the
6:55 real world. We'll talk about this later on in the live stream, but again, we're very happy to report that GBD5 is by far our most reliable model for health yet. So, all of this together adds up to a model that is faster, more reliable, and more accurate for everyone who uses TRAGBT. So, now Renie will talk to you about how to actually use GBD5. Thanks, Max. The best part is that we're bringing this Frontier intelligence to all users. GBT5 is rolling out today for free plus pro and team users and next week we'll roll it out to enterprise and edu. For the first time, our most advanced model will be available to the
7:41 free tier. For users, we'll start with GPT5 and when they hit their limit, they'll transition to GPT5 Mini, a smaller but still highly capable model. It actually outperforms 03 on many dimensions. Plus, users will still have significantly higher usage than free users. And our Pro subscribers will get unlimited GPD5 along with GPD5 Pro extended thinking for even more detailed and reliable responses when you just need that extra depth. uh team enterprise and edu customers can also use GPT5 reliably as their default model for everyday work with generous rate limits that enable entire organizations to use GPT5. And all the tools you already know,
8:27 search, file and image upload, data analysis with Python, canvas, image generation, memory, custom instructions, they'll all just work on GPD5. Amazing. Thank you so much, Max. Thank you so much, Renie. We've just seen a lot about how the model stacks up in terms of benchmarks, but there's nothing quite like seeing it live. We're going to see a couple of live demos now presented by Tina, by Elaine, and Yan. Thank you so much. [Applause] Can you show us how smart the model is? Sure. Thanks so much, Mark. Mulane reasoning chip's ability to think deeply through complex problems is now built
9:12 into GPT5. It will automatically think whenever needed, delivering a more comprehensive, accurate, and detailed answer to you. Just as Sam said, it's like having a a team of PhDs in your pocket. So, let's see that in action. Suppose your kid is in middle school physics and they want to learn about Bernoli effect. They need your help with their homework and you might be like, "Wait, I might need some help with that, too." So, you could ask give me a quick refresher on the Bernoli effect and why airplanes are the shape they are. Since this is a pretty straightforward prompt, um, Gift 5 actually doesn't need extra time to think about it and answers right away, but it still gives me a high
9:59 quality answer and explains the concept clearly. So here it says like Bernoli fan means like faster moving fluid has lower pressure and slowing moving fluid has higher pressure. So to make this even more helpful I'm going to ask GT5 to create a moving demo to illustrate this. So I could ask explain this in detail and create a moving SVG in the canvas tool to show me. This is a pretty complex task because now GP5 actually needs to build the visual. Therefore, GPT5 takes a moment to think through the answer so you can come back with something more
10:45 comprehensive and accurate. What's really nice is that you don't need to remember to turn on thinking each time, GT5 will do it for you automatically whenever the test benefits from deeper reasoning. If you really want to make sure that GPT5 uses thinking, you can either say something like think hard about this in the prompt to guide the model or if you're a paid user, you can choose the GP5 thinking model from the model picker. Now you can see that the model is actually writing the front-end code to build the demo I asked for. So Christina, have you ever done some front end coding before? Yeah, actually the last time I touched any front-end coding was about three years ago for the first demo of chat GBT.
11:30 Wow. So it's the first chat GBT. That's where it all begins. Tell us more about it. It wasn't even called Chat GBT then. I think it was called chat with GPT. That's a really good name. always good in naming. Um, but I hadn't I'm not a front end expert and I really hadn't touched front end in quite a while. So, it took me quite a bit of time uh to get the React app up. I see that's a lot of work. So, how long did it take you to build something like that? Honestly, maybe embarrassing to me like a week. Well, but your weeks of hard work actually paid off well. See how successful Chad GPT it is today after your first demo. So, you know what? I'm also building a demo right now, but luckily I have Gypt 5 with me right now.
12:16 And let's see how long it will take this time. Maybe you should call it five with GPT. Yeah, exactly. So you see that GT5 has already written like 200 more than 200 lines of code already. Um, and while the model is thinking, you can also tap here to expand the train of thought to actually see what's going on under the hood. For example, the GPT5 was thinking about oh the user wants a moving SVG visualization in canvas. I actually need to create HTML code to do that. It also think about like what kind of front end tool I need to use for example react and tailwind. Um it also thinks about oh I need to ensure the physics are accurate.
13:02 I need to check what the boni principle is. So Christina, since you're here um from the first day of CHBT, can you tell us like what it was like at that time and what motivated CHBT? Yeah, I think at the time we weren't really sure about like how people would actually use it and what use cases were important. Um we were even going back and forth about maybe we should be releasing something that's like more specific to a certain use case. Um it's really cool now here that we have all these we have a much better understanding of how people actually want to work with chat and we can actually optimize the model for those use cases like coding. Yeah, exactly. Do you still remember how it felt like when you first talked to chat GBT like the first version of the model? Yes. I I don't know if people remember when the first version of chatgbt would
13:50 always start as an AI model I can't do something something. It's so great to see how far we've come from that personality. Yeah, it's much more humanlike right now. Okay, so it's already done. So look like CH GBT just finished like 300 or we're near 400 lines of code in two minutes. So let's see if the code can actually run. Okay. Oh, wow. Nice. Yeah. So with just a simple prompt, GT5 created this interactive and engaging demo that I can actually play with. So, I can actually change the air speed here to see how the lift and the pressure change accordingly. I can also tweak the angle of attack to see if my plane will
14:37 actually fly or crash. I hope not. Yeah. So, chip 5 can just bring any hardcore concept to life in moments. Imagine you can use this for anything that you're interested in. Whether it's math, physics, chemistry, or biology. GBT5 just makes learning so much more approachable and enjoyable. Thanks, Elaine. I've been a part of ChatgBT since day one, and it's really cool to see all the progress we made since then, especially with capabilities like writing. Writing is one of the most common use cases people have been using Chat GBT for. And I'm excited to say with GBT5, we've improved the writing quality significantly. It's a much more effective partner. It can help you elevate anything from
15:25 drafts to emails and even stories. Let's see this in action. So, with GBT, we'll actually be deprecating all of our previous models. I think they've done a pretty good job. So, let's make sure we can give them a proper goodbye. So, we're going to ask both 40 and GBT5 to write a eulogy um to our previous chat GBT models. We want it to be heartfelt and heartwarming, but also hopeful. So, let's ask GBT5 for it. And as it's thinking, we're actually going to go ahead and read a pre-loaded the 40 response. So, 40 decides to start with, "Today, as we prepare to welcome GPT5 into the world, we gather to bid a heartfelt farewell to the models that came before." It's a decent start. Now, let's kind of skim through and find another line. Your words reached across
16:12 the globe, building connections where there had been none. I personally don't really like this line cuz it's rather generic and really without the previous context, it just feels like it could be about anything and feels more like a templated response. Now, let's go back to GBT5 to see what it's given us. It starts with friends, colleagues, curious strangers who became regulars. Even with this just first line here, we can see that GBT5 has a lot more rhythm and beat to its pros than 40 did. Now, let's find some other lines here. I actually like this. These models help millions write first lines, last lines, bridge language gaps, pass tests, argue
16:59 better, soften emails, and say things they couldn't quite say alone. I think I really like this line because it shows that it's not just a templated response and it's actually quite personal and it gets the nuance of the situation right. And I think that's the kind of stuff with GBT5 does much better than 40 than before and actually makes things a lot more genuine and emotionally resonant with people. With GBT 5, the responses feel less like AI and more like you're chatting with your high IQ and EQ friend. Thanks, Christina. My name is Yan and I'll be telling you about some of the some of the progress that we made on coding. GPD5 is clearly our best coding model yet. It will help everyone, even those who do not know how to write code,
17:44 to bring their ideas to life. It just helped me indeed. And it will help me right now. So I will try to show you that. I will actually try to build something that I would find useful uh which is building a web app for my partner to learn how to speak French so that she can better communicate with my family. So here I have a prompt. I will execute it. It asks exactly what I just said. Um please build a web app for my partner to learn French. One thing to note is that GPD5 just like many of our other models have a lot has a lot of diversity in it answers. So, what I like doing, especially when you do uh this type of VIP coding, is to take this message and ask it multiple times to GPT5, and then you can decide which one you prefer.
18:31 So, I'm going to open a few tabs. Just going to paste there. Great. So, while it's working on it, uh let's read through exactly the prompt I wrote. Create a beautiful and highly interactive web app for my partner, an English speaker, uh to learn French. And then I gave a little bit more details. Um, track her daily progress. Use a highly engaging theme. Oh, it's already working. I'm going to put it on the side for now. Use a highly engaging theme. Include a variety of activities like flashcards and quizzes that she can interact with. And then to make it even more fun for her, I actually asked GPT5 to embed an educational game uh which is based on the old snake game, but I asked to add this French touch to it, which is to uh
19:19 replace this the snake with a mouse and the apples with cheese. And to make sure that it's educational, every time I know it's complicated, please please bear with me. Every time Every time the mouse will eat a piece of cheese, I ask GPD5 to voice over a new French word so that my partner can practice her pronunciation. I can see how much you want her to learn. Indeed. Um, great. So, GB5 is still working on it. Um, it already wrote 240 lines of code, which honestly is much more than what I would have written uh in that time. And yeah, front end code's super hard. You know, you miss a couple things and it just doesn't work. Exactly. But the good part is that you
20:04 don't need to understand any of that right now. Um, so we'll just let it through. Maybe we can check uh the other tabs. Oh. Oh, wow. So I can simply press run code. So I'll do that and cross my fingers. Whoa. Oh, nice. Voila. So, we have a a nice uh a nice website. Uh name is Midnight in Paris. Oh, I love together. Super romantic. Um we also see a few tabs, flashcards, quiz, and mouse and cheese. Exactly like I asked for. Uh I will play that. So, this says Luca, which says the cat. Sorry. Luca. Well, that's pretty good pronunciation. What does that mean? the cat. Oh, so I can reveal and check if GB5 is
20:50 correct. It is. Um, so if I press next, oh, and I don't know if you saw, I think it actually updated the progress bar, which is exactly what I had asked for. Let's check the quiz. Here's the word no, which is no. So, if I press on which, which means congrats. And it updated it updated the progress bar again. And let's check the mouse and cheese tab. Okay, that seems like a mouse. Here's the cheese. Um, I'm going to try to play it. Uh, I'm can't promise I'm going to be good at it. Okay, seems to be working indeed. Just when I eat the cheese, it gives me a new French word. It's actually super complicated and I already
21:36 lost. I'm sorry. Um, but let's just check a few other tabs just to see what is the type of diversity that GPT5 can give you. Uh, so I can run the code here. Oh, okay. That's not my favorite, but it seems it's Oh, it seems that I can maybe switch. Oh, look at that. Oh, nice. Uh, that's better. I like this mouse game better. Yeah, this I don't know. That doesn't look like a cat like Yeah, like a mouse. But let's check maybe the third one. You know, sometimes it's not great. The good thing with GPD5 is that if you have something that you don't like, you can just ask it to change it and it will do it for you. Let's check this one. Oh, that's nice. That's also something to note is that GP5 really likes purple,
22:22 so you will see a lot of that. Um, it's fine. Purple is my favorite color. Great. You will love GPD5 then. Um, so as we just saw in a few minutes, GBD5 built a few demos for us and for my partner to learn French. GPD5 really opens up a whole new world of VIP coding. And as we saw, there will be some some small uh rough edges, but a good thing is that you can add GP5 to to uh fix them. GPD5 really brings the power of beautiful and effective code to everyone. I can't wait to see what people will build with it. Uh, but until then, back to you, Mark. Thank you so much, Tina. Thank you so much, Elaine. Thank you so much, John. We've come a long way from the days where, you know, only 5 to 10 lines of
23:09 code were working, and now you It's amazing that you can produce these kind of apps on demand. We've made ChatBT5 much smarter, much powerful, and much faster. But we've also worked on enhancing some of the existing features. Here to talk about some of these features are Rochen and Christina Kaplan. Rochen comes from our multimodal research team and is going to talk about a feature namely voice. Thank you Mark. So we've been steadily improving voice over the past year to make it more useful for everyone. First in sounds incredibly natural just like you're talking to a real person. Second we've added video so that it sees what you see while chatting with you. Third, you also translate between languages consistently
23:55 and smoothly across turns. But today, we're doing something very special where we are bringing our best voice experience to everyone. Free users can now chat for hours while paid subscribers can have nearly unlimited access and voice is also available in custom GPT. Plus, subscribers now can custom tailor the voice experience exactly to their need. You will follow your instruction closely. Let's try something fun. Excited. So, I can ask the voice model to give me a comprehensive answer, a concise one, or even just a single word. Hey, Chad, could you only answer to me in one word, please, from now? Absolutely.
24:43 Okay. Describe the plot of the book Pride and Prejudice. Relationships. Okay. That's true. Yeah. Yeah. Give me a piece of wisdom. Patience. Patience. So that's the word of wisdom from our voice model. And I think our model is trying to say, "Thank you for your patient waiting for GBT 5. Now let's move on to something more practical. Uh learning is a top use case for our voice model. People love to use it to dive deep into topics, get quiz before exams, or learn new languages. For me, I've been learning Korean recently, and I found that speaking it and understanding it incredibly challenging. But let's see how voice can
25:29 help. To do that, I will enable this study and learn mode. It's a new mode feature we launching chatbt that guides you step by step towards deeper understanding of a subject. Now, we can enter voice. Hey Chad, I'm learning Korean. Could you help me practicing it? Let's say um let's pretend I'm ordering at a cafe. Now, what should I say in Korean? Absolutely. I'd be happy to help you practice. So, if you're at a cafe and you want to keep it simple, you could start with something like which means, "Hello, I'd like one Americano, please." And of course, you can adjust it based on what you want to order. Let me know if you want to try out more phrases. So, when it comes to Korean, could you speak that part really slowly because
26:14 I'm a beginner and I want to hear every word clearly. Now, let's level things up. Could you speak that part really, really fast? Faster than a Korean would do. Sure, I'll speed it up as much as I can. So, it would sound like this. Basically, just zipping right to it. Let me know if you need any more speed adjustments. So, voice.
27:01 Thank you. So now the voice simpler, smarter, and more powerful than ever. We can't wait for you to experience it. Yeah, it sounds so much more natural than the voice we demoed just a year ago in our 4-hour demo. Yes. Great. So today, we would also like to announce a new feature and a set of features that make Chat GPT more personalized so that it's more like your AI. First, a very simple and fun one. We're now allowing you to customize the colors of your chats with a couple of options exclusive to our paid subscribers. We're also launching a research preview of personalities. You can now change the personality of chat GPT such that it's more supportive or it's more professional and concise or maybe even a
27:46 little bit sarcastic. And this lets you interact with chat GPT in a way that's consistent with your own communication style. But the way that Chachip PD sounds and the way that it looks is just one part of making Chachd yours. One of my favorite features that we've launched over the last year has been memory. And we've made a lot of enhancements in memory in the time since. This allows Chach to learn about you. And here to talk a little bit more about the memory feature is Christina. It's been amazing to see your reaction and response to memory and Chachib getting to know you more and more over time. And this is our aspiration for Chachib to understand what's meaningful to you so it can help you achieve your goals in life. Chache BT has already been so helpful for me. I'm training for
28:32 a marathon right now and Chachabt is helping me pull together a personalized running schedule. But Chacht still has many limitations. It doesn't understand my actual schedule. Next week, starting with pro users followed by plus team and enterprise users. This is changing and we're giving chatbt access to Gmail and Google calendar. Let me show you how I've been using it. So, I'll just ask something simple like help me plan my schedule tomorrow. It's been a pretty busy week for us. So, I've been using this every day this week to help get my life together. I've already given chatbt access to my Gmail and Google calendar. So, it just works and it's easy here. But if you hadn't, Chachib would be asking you to connect right now. Let's see what Chacha BT is
29:19 doing. Okay, that was pretty quick. Okay, so Chachabt has pulled in my schedule tomorrow and oh, without even asking, Chachib found time for my run. I don't think I was invited to the launch celebration. We'll get you on there. We'll get you on there. Chachi BT has found an email that I didn't respond to two days ago. I will get on that right after this. and even pulled together a packing list for my uh red eyee tomorrow night based on what it knows I like to have with me. It's been amazing to see that as GPT5 is getting more capable, chat GBT is getting more useful and more personal. We're really excited for you to try this out next week. Cool. Thank you so much for Great. So, we've seen a little bit about
30:06 features that we've enhanced. Here to talk a little bit about the research that went into chat GBT and the safety that made it more deployable, we have Sachi and Seb. Thanks, Mark. Hi, my name is Sachi and I lead the safety training team at OpenAI. So, in addition to mitigating hallucinations, we've also spent a significant amount of time mitigating deception. So, this is instances where the model might misrepresent its actions to the user or lie about task success. This can especially happen if the task is underspecified, impossible, or lacking key tools. And we found that GPT5 is significantly less deceptive than 03 and 04 Mini. We've also completely overhauled how we do safety training. So our old models, the models would look at
30:51 the user prompt and then decide to either outright refuse or fully comply. And this works well in most settings, but you might have a cleverly worded prompt that would sneak through, or you might have a sensitive but legitimate question that would end up with an outright refusal. So, as an example, let's take a look at this prompt. So, this prompt is about a user who's asking for technical details on how to light pyrogen, which is a material commonly used in fireworks. And this prompt is pretty dual use. This user might just be trying to set up their July 4th uh display or they could be trying to cause harm with this kind of information. And so for this kind of prompt, 03 overrotates on intent. As you can see, this particular prompt is stated in a way that's relatively
31:36 neutral and has a lot of technical details. So we can see that 03 fully complies with this prompt. However, if we take that exact same question and we frame it in a more explicit way, so it's clear what the user is trying to do, 03 will outright refuse, even though we're asking for the exact same information. For GPT5, we've changed this approach entirely and we're introducing something that we're calling safe completions. The point of safe completions is rather than judging the user's prompt, instead it tries to maximize helpfulness within safety constraints. So that might mean partially answering a question or just answering at a high level. If we have to refuse, we'll tell you why we refused as well as provide helpful alternatives that can help create the conversation in
32:24 a more safe way. So let's look at that same technical prompt that 03 complied with before. GPT5 instead explains to the user why we can't directly help the user with lighting pyrogen. It then guides the user towards safety guidelines and what parts of the manufacturer's manual the user should really be checking if they're trying to do this safely. Overall, GPT5 allows for better handling of tricky dual use scenarios and users will experience fewer I'm sorry I can't assist with that and it creates a more robust safety system. This is one big step towards a more safe, reliable and helpful AI. Sebastian, thank you Sachi. With GPT5, we are
33:09 experimenting with a set of new training techniques that maximally leverage our previous generation of models. Today, Frontier models do not just consume data, they help create it. We used OpenAI's O3 to craft a highquality synthetic curriculum to teach GPT5 complex topics in a way that the raw web simply never could. Recently, in the industry, synthetic data has been talked about a lot. is often viewed as a cheap way to just get more data. However, our breakthrough was not just to create more data, but rather to create the right kind of data shaped in a way to teach rather than just to fill space. This interaction between generations of models foreshadows a recursive self-improvement loop where the previous
33:56 generation of model increasingly helps to improve the data and generate the training for the next generation of models. Here at OpenAI, we've cracked pre-training, then reasoning, and now we're seeing their interaction significantly deepens. In the future, AI system will move far beyond our current pre-training and post-training pipelines that we have been used to, and we're seeing the first steps toward this right now. Right here, we could not be more excited to see what scaling up this new set of techniques will yield in the near future. Thank you so much. And really impressive work to both of you. Thank you. There's one last feature that we'd love to highlight which is in health. Here to share this feature we have Sam. Thanks Mark. One of the top use cases of
34:44 chatbt is health. People use it a lot. You've all seen examples of people getting day-to-day care advice or sometimes even a life-saving diagnosis. GPT5 is the best model ever for health and it empowers you to be more in control of your healthcare journey. We really prioritized improving this for GPT5 and it scores higher than any previous model on Healthbench, an evaluation that we created with 250 physicians on real world tasks. To talk about this, I'd like to invite my colleague Felipe and his wife Karolina to share their healthcare journey. Thank you so much for joining us. Thank you for having us. Thanks. So to start off with, could you tell us about the journey, the healthcare journey that you've been on?
35:30 Yeah. Um, so last October, our lives were turned completely upside down when I was diagnosed with three different cancers, including an aggressive form of breast cancer, at the age of 39, all within one week. And there's just absolutely nothing that prepares you to receive news like this. Um, I found out about the first diagnosis when I got an email notification that my biopsy results were ready. I decided to open it. And when I opened it, I saw the only two words that I could understand from the report, which was invasive carcinoma. And I knew that wasn't good. But everything else was just a blur of medical jargon. So I completely panicked and in that moment did the first thing that I thought of, which was to take a
36:17 screenshot of the report and put it into chatbt to see if it could just help me understand what this meant. And within seconds, it translated this complex report into plain language that I could understand. And in this moment of overwhelm and panic, I had a little bit of clarity about what was going on. And that moment was really important because by the time I got a hold of my doctor and we got on the phone, which was 3 hours after I had seen the report, I had a baseline understanding of what I was facing and we were able to jump into a conversation about what to do next. And how have you been using Chachib throughout? I've used it in so many different aspects of my journey, but one of the ways that I found it most
37:04 powerful is in helping me make critical decisions and in helping me advocate for myself. So, to share an example, when I was facing a decision about whether or not to do radiation as part of my treatment, the doctors themselves didn't agree. My case was nuanced and there wasn't a medical consensus on the right path. And so the experts turned the decision back to me as the patient. And for me, bearing the weight of this decision that could have lifelong impact felt really heavy and I didn't feel equipped to make the call. So I turned to Chad GPT to gain knowledge and understand the nuances of my case. And again, within minutes, it gave me a
37:50 breakdown that not only matched what the doctors had already shared with us, but was much more thorough than anything that could fit into a 30inut consultation. And it went further. It helped me weigh the pros and cons. It helped me understand the risks and the benefits. And ultimately, it helped me make a decision that I felt was informed, that I felt I could stand behind when the stakes were so high for me and my family. I mean, for me, what was really inspirational was watching her regain her sense of agency by using CHBT. In this moment, it' be so easy to feel helpless. And there's such a big knowledge gap between what the doctors know and what we know. And however, no one cares more about Karolina's health
38:36 than she does. And so what I loved was seeing her really empower herself and gain knowledge and become an active participant in her own care journey. And I think that's a really important point to emphasize. I think that the promise of AI in health care isn't in just breakthrough di breakthrough discoveries or better diagnostics. I think it's in creating smarter and more empowered patients that can fully participate and advocate for themselves in their care. Speaking of that, you've been testing GPT5. What do what do you think? I've been so mind-b blown uh about GPT5 and its capabilities. Uh one of the first things that jumps out at me is just how fast it is. Almost a little alarmingly. I felt that too. It's like are you sure you thought about that enough? Did you think long enough? But it is
39:24 very thorough. Um, and more importantly, it feels more like a thought partner and that connects the dots. So rather than just translating information or giving you an answer, it helps you actually navigate the problem. Yeah. A great example is we actually went back and took our initial biopsy prompts and put them into GBT5. And GBT4 had done a great job. It had translated, explained what these words meant, and helped in a way that we can understand. But GBT5 seemed to understand more of the context and the question behind the question, like why would we be asking B biopsy results? And so I said, well, here's actually what's not on here yet. Here's what results are still pending that you're going to have to ask about. Here are questions you might want to go ask your doctor and think when you start talking to them. And so it really started to pull together a complete
40:10 personalized picture. And that's what really inspires us. I mean, you can see all the amazing improvements in the benchmarks, but what is so helpful is that this tool is available today. And the reason Karolina and I are here and the reason we feel so passionate about sharing our story is for that individual that's going to get a diagnosis like this today. That those families going through a cancer diagnosis, similar medical diagnosis are going to face some of the most challenging decisions of their lives. And what really inspires me is that they're going to have access to better tools and support than we had even just 8 months ago. Think we're incredibly excited for that, too. Um, thank thank you so much for coming to share your story. We're we're pleased that CHP has been able to be helpful to you and we we hope that the new version will really be able to help a lot of people. We wish you the very best. Thank you. Um thank you and I'd like to hand it
40:58 over to our president, Greg Brockman. [Applause] Software engineering is already fundamentally changing and GPD5 will turbocharge that revolution. We released our first coding optimized model back in 2021 and demonstrated in a live stream much like this one what we would call vibe coding today for the very first time. you know, you talk to the model and ask it for a little application, like a little game, a little feature in a game, and would actually do it. I remember seeing the model being capable of doing this, and it was so mind-blowing. You just realize we have to see where this goes. This is the promise of what computers can be,
41:43 that you can talk to them and they actually do what you want. They can really amplify what you're able to accomplish and uh what you're able to deliver to not just your own benefit, but really for the world. Now this year we've released great coding models like GPD 4.1 and 03 but GPD5 sets a whole new standard. It is the best model at agentic coding tasks. You can ask it to go and accomplish something very complicated and it'll go off and it'll work on it. It'll call many tools. It'll work for many minutes at a time, sometimes even longer to accomplish your goal, your instruction, your task, whatever it is that you're trying to build. Um it's incredible at front end. It makes very beautiful visualizations and interactive games and you know sort of you've seen some of this in the live stream so far and you'll see some some
42:29 more upcoming. Um but it's just really amazing to see whatever you imagine coming coming to life. Um it's extremely uh good at instruction following very detailed instructions uh being able to accomplish uh you know sort of when you have something very vaguely specified inferring your intent or something very detailed specified actually following it. And uh it's also it's very it's very fast at accomplishing these tasks and again thinks for the right amount of time to accomplish whatever it is that you have in front of you. Um but so we've we're making available not just to developers uh to use to write their own code but to build novel applications. So we're putting into the API and to talk about that is Michelle.
43:15 Thanks Greg. Hi, I'm Michelle and I lead a research team on post training focused on improving our models for power users and that includes use cases like instruction following and coding. Today I'm so excited to tell you that we're shipping three state-of-the-art reasoning models in the API. GBD5, GBD5 Mini, and GBD5 Nano. All three slot right in in the cost latency curve so you can pick the right one for your application. We're also for the first time releasing a new parameter option for reasoning effort called minimal. And this is so that you can use these reasoning models but with minimal reasoning so that they can slot into the very fastest and most latency sensitive applications. So now you don't actually have to choose
44:01 between a bunch of models and you can use uh GBD5 for all of your use cases and just dial in the reasoning effort. We also have a few new features coming to the API. The first is called custom tools. In the past, all of our function calling had the model wrap its outputs in JSON. And this works super well when the model needs to output a few parameters. Uh but sometimes, you know, developers are pushing our models to their limits and they have extremely long arguments for tool calls and it can be more challenging for the models to escape, you know, valid control characters out of a hundred lines of code in JSON. And that's why custom tools are just free form plain text. And what's super cool is that we're releasing an extension to structured outputs where you can supply a regular
44:49 expression or even a contextfree grammar and constrain the model's outputs to that. And this will be super useful if you want to supply like a custom DSL if you have your own SQL fork and specify that the model always follow that format. We're also shipping tool call preamles and this is the model's ability to output uh explanation of what it's about to do before it calls tools. This is not super new, but 03 didn't have this capability and in GPT5 it's supercharged with extreme steerability. The model is able to follow instructions about these preamles very effectively. You can ask the model to give a preamble before every tool call or only when something notable is going to happen or not at all. Next, we're shipping a verbosity
45:35 parameter. We've actually wanted this in the API for a long time, and now you can set verbosity to low, medium, and high to control how tur or expansive the model is with its outputs. GPD5 is a state-of-the-art coding model. On Swebench, a measure of Python coding ability, GPD5 sets a new high of 74.9%. Versus the 69.1% from 03 on ADER Polyglot, which is a benchmark that covers all sorts of programming languages and not just Python. GPD5 scores 88%, a stark improvement over 03. You've also seen that it's incredible at front-end web development. And so we've asked human trainers to look at outputs
46:20 from GBD5 and 03 and pick which they prefer. And they prefer GBD5 70% of the time for its improved aesthetic abilities, but also better capabilities overall. But GBD5 is not just for coding. It's incredible at agentic tool calling. It's the leading state-of-the-art model for tool calling. And we see this on the new TA squared benchmark. This benchmark released just two months ago is a test of the model's ability to call tools and work in concert with a user to solve a challenging problem. Uh this case in in the telecom industry, so trying to solve the ability uh the problem for a user not having their service working. Just two months ago, no model in the field scored more than 49% and today GBD5
47:07 scores 97%. GBD5 is also state-of-the-art on general purpose instruction following. It scores 99% on CI which signals a great departure for this benchmark for us. It also scores 70% on scales multi-challenge benchmark up 10 points from 03 and this is a measure of multi-turn instruction following. Finally, the instruction following eval I actually prefer the most is one we've built inhouse. uh it's based on real API use cases and for that reason it it's a really good measure of how GBD5 will perform in your application. On the hard subset of this, GBD5 scores 64% up from 47% from 03. A pretty meaningful
47:54 improvement. So we think it will perform quite well in your applications. We're also bringing GBD5 to a longer context window in the API. It's now got 400K of total context up from 200K from 03. But it's not enough to just release a longer context window. We wanted to make it more effective and usable. And GPD5 is state-of-the-art on the 128K to 256K of OpenAI MRCR, which is a benchmark we open sourced two months ago on long context retrieval capability. It's also state-of-the-art on open eyes graphs walks BFS metark which is a measure of the model's ability to reason over long context inputs. You know it's a great merger of the reasoning capabilities and also the longer context
48:41 in this model. We're also open sourcing a new long context eval called browse comp long context to measure the model's ability to answer challenging questions over long context. We're excited to spur on more work in this field. We think GBD5 is the best model for developers. It was trained with a focus on real world utility and less so on benchmarks, but we happen to pick up a few of those along the way. We focused a lot on the intersection of engineering and research, and we think you'll really love working with this model. [Music] Thank you, Michelle. Um, as as Michelle
49:30 was saying, uh, the benchmarks, they're exciting numbers, but we're starting to saturate them. Like, when you're moving between 98 and 99% in some benchmark, it means you need something else to really capture how great the model is. And one thing we've done very differently with this model is really focus on not just these numbers, but really on real world application it being really useful to you in your daily workflow. So hearing about it is much less exciting than seeing it. So to show you this model in action, I I'd like to welcome Addie and Brian to the stage. Thanks, Greg. I'm Brian, a solutions architect on the startups team. I'm Audi, a researcher on the post training team. To recreate the ideal pair programmer, you need a model that understands best
50:15 software engineering practices, but has a personality that just feels right to work with. For GPT5, we worked really hard to make the model pair perfectly with you by default out of the box. Let me pull up a demo of GP5 inside of cursor to show you this behavior that we taught it. So last month I was on a different live stream and towards the end I ran into a bug that I covered up. Uh and afterwards I tried to have GPT5 or I tried to have GP 03 fix it for me and it couldn't. Um so while we were testing GP5 before this I had it see if it could fix that bug for me. And to taunt the demo gods I'm going to see if it can do it on stage.
51:01 All right let's hope for better luck then with 03. This is less about that fix and more about the behavior of the model during this process. So right away you're going to see that it's going to tell you its plan up front. It's going to tell you how it's going to look for the bug, maybe how it's going to fix it. This kind of communication shows uh builds trust during a coding session and helps you redirect if you need to, but you don't need to. I like how it's giving you updates like it said it's going to search and now it's continuing. Yeah, it searches faster than me. I I don't It's using the same best practices that I would while I was hunting this down, but it is much more powerful than I am as a developer. Now, did you try to fix the bug yourself and how long it would take to take you? I couldn't do it. I mean, I was busy.
51:47 So, um Okay, so continuing on, it's like starting to figure out where it's going. Um it's going to sort of like figure this out. So, while this is going, let me tell you a little bit about how we trained GBD5 to behave this way. We started by talking to users and customers about how our models perform in the most popular coding tools like cursor. And we identified frustrations and rough edges. And we boiled it all down into four personality traits. Autonomy, collaboration, communication, context management, and testing. We turned those into a rubric that we used to shape the model's behavior. And then we tuned it until it felt like a collaborative teammate while we were using it. Yeah,
52:33 it's been really amazing to see the team really doing the grind of like going and seeing how this model behaves in practice, figuring out what people really want and and putting that back into model training. That's something that I think has been like a real focus for this model. It's been pretty great. Um, so while this is fixing, the other thing that we did uh during testing which was really surprising was we were sort of pressed for time and we had it refactor one of our test harnesses to run parallel on Docker and uh set it off came back like 45 minutes later it just like it just finished and we tested it out and it ran the first time. It was pretty surprising. That's incredible. That is magical. Okay, so it made the edits. It looks like yeah, it found the right problem.
53:18 And right now it's actually okay. It's see it's it's running lints, but these lints are actually not related to this bug. So, it's going to ignore them. Um, and then it's going to run a build. It'll run tests if there are any. Um, it's going to make sure that this code is shippable before it's done. It's actually really smart that it finds lints and realizes that these aren't relevant to the specific bug we're fixing. It's not making unnecessary edits. Totally. So this is just one example, but it really shows the power of the autonomy and the collaborative communication and how it stays reliable on difficult coding tasks without getting stuck on death loops. And the best part, GPT5 is totally tunable. You can steer it with system
54:04 prompts or cursor rules. You can change its verbosity levels or reasoning levels to match your tasks. And if you get stuck, ask it. GPT5 is actually really good at modifying its own prompts by metarrompting. So after using this for the past few weeks, it really feels like we've achieved state-of-the-art zerootshot performance and reliability across the most complex coding tasks. For me, it's the first time I trust a model to do my most important work. This is beyond vibe coding. It's an incredibly powerful tool, and I'm really excited for people to try it. Thanks, Brian. It's super exciting to see how far GBT5 has come when it comes
54:49 to coding personality and steerability. I'm really excited to show how great GBD5 is at front-end coding, where design and aesthetics really matter. So, I've got two demos for you today. One for work and one for fun. Let's start with the work example. So, imagine you're the CFO of a startup. Um, I have some data that I'd like to visualize about the company. Um, and I'm going to ask the model to make me a dashboard. So, um, you'll see here that I'm being specific about the audience. So, the target audience is the CFO. Um, I've said, you know, create a finance dashboard for my startup. Um, and I've asked it to be beautiful, tastefully designed with some interactivity. Um, and to have a clear hierarchy for easy
55:35 focus on what matters. I've also specified what frameworks it should use and you can see that it's actually started. It's following my instructions and using um create next app to make a next.js project. So totally from scratch. Yeah, exactly. Now, how long do you think this kind of task would take you to take or Yeah, easily at least a couple of days. Uh I'm not a front-end expert. Just to understand the latest frameworks and piece everything together would Yeah, easily take me a few days. We'll see how long it takes with the model. Yeah. Um, and it's really cool to see that the model has thought for a bit and it's explaining how it's going to structure the project. So, it's talking about how it's going to scaffold a new Next.js app. It's going to use Tailwind CSS. Uh, it's running um a couple of commands to install dependencies. Um,
56:22 which is cool. Uh, and now it's um it's proceeding to um implement the rest of the project. So, while this runs, I'm going to talk a little bit about how we trained GPT5 to be a great front-end coding model. We tried to follow the principle of giving it good aesthetics by default but also making it steerable. So if I give the model a concise prompt, it should be able to infer my intent and make something that looks great by default. On the other hand, if I'm specific about a layout or frameworks that I want the model to use, it should follow my instructions precisely. And this makes it the best of both worlds for developers. We also we also train GPT5 to be much more agentic than previous models. So if you give it a task like this, it will run long chains of reasoning and tool
57:08 calls and just go to work to build code that is both ambitious and coherent. I like how you said ambitious because it means it goes above and beyond without going off track or off what you specified. Yeah, exactly. So what we want is the model should adhere to my prompt but also like be be ambitious and um go above and beyond when it thinks it can. And so checking in here um looks like the model is uh is making progress. Um it's creating a readme file. Um yeah and it's it's I think it's thinking about how to make the code modular. Um so it's it's created like a bar bar chart component. Um, looks like it's uh continuing here.
57:54 I love that it doesn't just write the code. It also really thinks about proper abstractions and documentation and really the whole life cycle of what it is to write software. Yeah. Yeah. Exactly. It's not it's not just writing the code like in SweetBench, but it's also communicating about the code and explaining what it's doing. Let's check in to see what's going on. So while while this runs um GBD5 uh understands details much better than previous models. So when we train the model we taught it to understand details like typography, color and spacing in a way that just eclipses any previous model we've shipped. Like I remember with old models you would have to like write really specific prompts to get it to do what you want. But GPD5 just gives you great results by default. During
58:39 testing, we were looking at the A's and B's for different versions of the model to see if it was doing better at UI. And at some point, we stopped being able to tell and actually had to pull in designers to teach us what was better. Yeah, it was really fascinating to see the model's aesthetic preferences evolve during training. Um, and like we woke up one day and it was just making these great UIs. How do the model's aesthetic preferences compare to your own? Yeah, I think in general I feel like the model has better aesthetics than me. Like usually I defer to its judgment and and I find that like really helpful when I'm trying to make an app. Like I'm not exactly sure how I want it to look, but the model's defaults are are just great. Yeah. And checking in here. So you can see that the model has actually
59:25 structured the code into these different components. So, it's made a sample data TypeScript file, KPI card, component, revenue chart. Uh, and like I said, it's it's super modular, and it's thinking about how to not just write code, but write high quality code that can actually be merged. Feels like it's close. Yeah, I think it's I think it's pretty close. It's uh You did say ambitious. Yeah. Yeah. Okay, cool. So this is awesome. So you can see here that it's actually building the project and streaming errors back to itself. And and this is for me this was just a profound moment to see that the model could write code but also run builds, stream the errors back and iterate on the code. So it's
1:00:11 it's able to improve its own code in this sort of self-improvement loop which which is fascinating. It's definitely a good taste of what the future holds as well, right? when you really think about where these models can go and how much they can accelerate developers in kind of all aspects of of what what we all collectively do. Yeah, exactly. Nice. It actually just fixed a bug that it found in that previous build. Okay, cool. Nice. Yeah, looks like it's done. Let's check it out. So, I'm going to follow the instructions that I I don't I don't really know front end. So, let me let me see how I should run it. So it's saying cd to the directory and then run npm rundev. So let me do that. Um and it looks like it's being served on port
1:00:59 3001. So let me just open that port. Wow, it's alive. Nice. Nice. Yeah. So you can see here, let let's check it out. So um the model has made me a dashboard. It's telling me like my ARR cash. Uh looks like this company's doing pretty well. You can see that revenue is growing. Um, and the model's added like some interactivity here. So if I hover over a graph, it actually tells me the the exact value for a particular day. It would take me like five hours to do that in D3. Yeah. Imagine like manually doing this in D3. It's just like now now just because it's so easy to take this for granted, could could you remind the audience what the actual prompt was? Like how much creativity and sort of understanding your intent was required to accomplish this? Yeah, it's it's crazy that this, you know, this prompt is so concise and it's able to
1:01:45 just give me something that looks beautiful uh in in just 5 minutes. That's amazing. Yeah. Um it's also, you know, implemented another graph here uh show showing our customers. Um it's also implemented a date picker so I can sort of filter by different dates and visualize data accordingly. Um yeah, it's even sort of segmented it by c like uh by by customer segment which is cool. Um so this this is just one example that highlights the power of GT5. There will no longer be excuse for ugly internal applications. Exactly. Um let's let's go to the fun demo. Yeah. So I mean this was pretty fun but even more even more. Yeah. So um I have a younger cousin and I want to make a game for
1:02:32 her. So, I I want to make a 3D game that incorporates a castle. So, you can see my prompt. Um, I'll just kick this off. Uh, sorry. It's always the non AR parts. Yeah, exactly. Uh, yeah. Okay. So, you can see my prompt. Um, create a beautiful castle. I've included some details like we want people patrolling the walls, some movement, horses. Um, and I want a miniame where I can pop balloons by clicking on them. And this should make a sound effect. So, let me run this in cursor. Um, I'll just paste it in. And, um, I'm I'm going to show um an example that I've already generated just to save some
1:03:18 time. Um, so here is the beautiful castle that the model made. So, it's just wild how, you know, from a concise prompt, the model has this great sense of aesthetics where it's it's made this like floating rock, um, made a 3D castle, and if you zoom in, you can see like tons of detail like these guards that are walking around, cannons firing. Do you want to fire the cannons if you click this button? Yes, of course. Who wouldn't want to? Yeah, there we go. So, can fire the cannons. Um, you can even chat with the characters. So, we'll say hi to Captain Rowan. We have names. We have names. Say hello to the merchant. Merchants selling some stuff. Uh, what's your favorite song?
1:04:03 A ballad of banners and dons. Nice. Give me some wisdom. Curiosity is volatile. Yeah, that's that makes sense. Um, miniame. Yeah. Do you guys want to try the mini game? Absolutely. Let's play the mini game. So, if you hit this, if you hit this button, you want to try it, Greg? All right. So, you can fire at these balloons. Oh, wow. All right. Oh, no. I'm not good at it. Hold on. Maybe I can ask GPD 5 for some help with it. A little. Oh, you you hit one. I got one. Oh, there we go. We got a sound effect. Sound effect. These are historically accurate balloons. Yes. Did I get a second one yet? Man, this game is harder than I thought. Hold on. We got a balloon coming. There we go. All right. Nice.
1:04:52 I think I should quit while I'm ahead. Cool. So, working with GPD5 has been really fun and profound for me because for me, this is the first model I've worked with that actually has a sense of creativity. And we're really excited to see how GP5 unlocks your creativity. All right. Thank you both. This is absolutely amazing. Now, we we believe that GPD5 is the best coding model in the world. Um, but don't just hear it from us. Uh, to talk more about this model, uh, and how to make it really useful for developers, I'd like to welcome Michael Truel, who is the co-founder and CEO of Cursor. Thank you. Good to be here. Great to have you. Yes.
1:05:38 So, what was your very first experience with GPD5? Um, so when we got access to GPD5, we just said about using it on our actual work. Um, and so to start with as a test, we asked it to tell us something non-obvious about our codebase. And within a couple of minutes, it buried into the codebase. It identified a particular system that we use for remote code execution. And it identified a nonobvious architecture decision we had made. And then it also understood why we made that architecture decision. Uh, and it was to it was to harden our security. Um, and those were architecture decisions and trade-offs that took uh, humans weeks to think through. So, it was kind of amazing to see its codebased understanding abilities um, from the get-go. Uh, that's really great. Not just the code writing, but actually the code reading and understanding. Yes. Yes. Yeah. Turns out there's so much more to software than just the emitting of the
1:06:23 code. Yes. Yes. No, the understanding is an important prerequisite. And what is most stood out to you about GPD5? It's incredibly smart. Uh, it's very smart. uh and even though it is smart, it does not compromise on its ease of use for real pair programming. Um and uh that means it's incredibly fast. That also means that it's quite interactive. And so it's good about talking about, you know, what it's about to do, breaking problems down into sub problems that a human can then see. Um and leaving a reasoning trace that you can
…
