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 them. 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 to 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 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 ChatGPT 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 enjoy 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.
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 chat agent and deep research. GPT5 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 GPT5, it eliminates this choice. It aims 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 GPT5 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 GPT5 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 GPT5 supercharges our ChatGPT app and our API. We believe that GPT5 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 GPT5 is by far our smartest model ever. So, let's start by talking through some evals. Now, evals aren't everything and they don't tell you everything about a model, but they can highlight its intelligence. And GPT5 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, GPT5 sets a new high on SWE-bench, which is an academic eval that tracks performance on real software engineering tasks. Now this again is an eval but we think it will reflect the model's performance in the real world. GPT5 also performs very well on Aider polyglot which measures its ability to implement complex functionality in a variety of different programming languages.
Now, beyond coding, GPT5 performs exceptionally well at multimodal reasoning, setting a new high on MMMU, 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. GPT5 is also excellent at mathematical reasoning as shown by its performance on AIME 2025. Now this is an exam that American high school students take to qualify for the international mathematical olympiad and GPT5 performs exceptionally well again beating our previous models and other models that are out there.
Now moving beyond academic evals towards some real world use cases. We put a lot of work into making GPT5 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 GPT5, 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 GPT5 is by far our most reliable, most factual model ever. GPT5 also performs exceptionally well on health related questions. Now, health is a big part of how people get value from GPT in the real world. We'll talk about this later on in the live stream, but again, we're very happy to report that GPT5 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 ChatGPT. So, now Renie will talk to you about how to actually use GPT5.
Thanks, Max. The best part is that we're bringing this frontier intelligence to all users. GPT5 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 free tier. Free users will 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 o3 on many dimensions.
Plus, users will still have significantly higher usage than free users. And our pro subscribers will get unlimited GPT5 along with GPT5 Pro extended thinking for even more detailed and reliable responses when you just need that extra depth. 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, search, file and image upload, data analysis with Python, canvas, image generation, memory, custom instructions, they'll all just work on GPT5.
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.
Lane, can you show us how smart the model is?
Sure. Thanks so much, Mark. Mulane reasoning ChatGPT's ability to think deeply through complex problems is now built 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 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 the Bernoulli 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 Bernoulli effect and why airplanes are the shape they are.
Since this is a pretty straightforward prompt, GPT5 actually doesn't need extra time to think about it and answers right away, but it still gives me a high quality answer and explains the concept clearly. So here it says the Bernoulli effect means faster moving fluid has lower pressure and slower moving fluid has higher pressure. So to make this even more helpful I'm going to ask GPT5 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 GPT5 actually needs to build a visual. Therefore, GPT5 takes a moment to think through the answer so it can come back with something more comprehensive and accurate. What's really nice is that you don't need to remember to turn on thinking each time. GPT5 will do it for you automatically whenever the task 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 GPT5 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 ChatGPT.
Wow. So it's the first ChatGPT. That's where it all begins. Tell us more about it.
It wasn't even called ChatGPT then. I think it was called Chat with GPT.
That's a really good name. We're not always good at naming. But 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 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 ChatGPT is today after your first demo. So, you know what? I'm also building a demo right now, but luckily I have GPT5 with me right now. 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 GPT5 has already written more than 200 lines of code already. 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 thinks about like what kind of front-end tools I need to use, for example React and Tailwind. It also thinks about, 'Oh I need to ensure the facts are accurate. I need to check what the Bernoulli principle is.'
So Christina, since you're here from the first day of ChatGBT, can you tell us like what it was like at that time and what motivated ChatGBT?
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. We were even going back and forth about maybe we should be releasing something that's more specific to a certain use case. It's really cool now here that 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 ChatGBT, like the first version of the model?
Yes. I don't know if people remember when the first version of ChatGPT would always start with, '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, ChatGPT 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.
Nice. So with just a simple prompt, GPT5 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 actually fly or crash.
Yeah. So, GPT5 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. GPT5 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 ChatGBT for. And we're excited to say with GPT5, we've improved the writing quality significantly. It's a much more effective partner. It can help you elevate anything from drafts to emails and even stories.
Let's see this in action. So, with GPT5, 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 4o and GPT5 to write a eulogy to our previous ChatGBT models. We want it to be heartfelt and heartwarming, but also hopeful. So, let's ask GPT5 for it. And as it's thinking, we're actually going to go ahead and read a pre-loaded 4o response.
So, 4o 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 the globe, building connections where there had been none.' I personally don't really like this line because 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 GPT5 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 GPT5 has a lot more rhythm and beat to its prose than 4o did. Now, let's find some other lines here.
I actually like this: 'These models helped millions write first lines, last lines, bridge language gaps, pass tests, argue 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 that GPT5 does much better than 4o did before. It actually makes things a lot more genuine and emotionally resonant with people. With GPT5, 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 progress that we made on coding. GPT5 is clearly our best coding model yet. It will help everyone, even those who do not know how to write code, to bring the ideas to life.
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, 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: please build a web app for my partner to learn French.
One thing to note is that GPT5, just like many of our other models, has a lot of diversity in its senses. So, what I like doing, especially when you do this type of vibe coding, is to take this message and ask it multiple times to GPT5 and then you can decide which one you prefer. So, I'm going to open a few tabs. Just going to paste there. Great. So, while it's working on it, let's read through exactly the prompt I wrote.
Create a beautiful and highly interactive web app for my partner, an English speaker, to learn French. And then I give a little bit more details: 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 which is based on the old snake game, but I asked to add this French touch to it, which is to replace the snake with a mouse and the apples with cheese. And to make sure that it's educational, every time the mouse eats a piece of cheese, I ask GPT5 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. Great. So, GPT5 is still working on it. It already wrote 240 lines of code, which honestly is much more than what I would have written in that time.
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 don't need to understand any of that right now. So we'll just let it through. Maybe we can check the other tabs. Oh.
So I can simply press run code. So I'll do that and cross my fingers. Whoa.
So, we have a nice website. Name is Midnight in Paris together.
We also see a few tabs, flashcards, quiz, and mouse and cheese. Exactly like I asked for. I will play that. So, this says 'le chat,' which is 'the cat.' Sorry. 'Le chat.'
Well, that's pretty good pronunciation. What does that mean? The cat.
Oh, so I can reveal and check if GPT5 is correct. It is. 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 'non,' which is 'no.' So, if I press on it, which means congrats. And 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. I'm going to try to play it. I 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 lost.
I'm sorry. But let's just check a few other tabs just to see what is the type of diversity that GPT5 can give you. 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.
I like this mouse game better.
Yeah, this I don't know. That doesn't look like a cat, like a mouse. But let's check maybe the third one. You know, sometimes it's not great. The good thing with GPT5 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 GPT5 really likes purple, so you will see a lot of that.
It's fine. Purple is my favorite color.
Great. You will love GPT5 then. So as we just saw in a few minutes, GPT5 built a few demos for us and for my partner to learn French. GPT5 really opens up a whole new world of vibe coding.
And as we saw, there will be some small rough edges, but a good thing is that you can add GPT5 to fix them. GPT5 really brings the power of beautiful and effective code to everyone. I can't wait to see what people will build with it. But until then, back to you, Mark.
Thank you so much, Tina. Thank you so much, Elaine. Thank you so much, Yan. We've come a long way from the days where, you know, only 5 to 10 lines of code were working. And now it's amazing that you can produce these kind of apps on demand. We've made ChatGPT much smarter, much more 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, it sounds incredibly natural, just like you're talking to a real person. Second, we've added a video so that it sees what you see while chatting with you. Third, you can also translate between languages consistently 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. It will follow your instruction closely. Let's try something fun.
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 for now?
Absolutely.
Okay. Describe the plot of the book Pride and Prejudice.
Relationships.
Okay. That's true.
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 GPT5.
Now let's move on to something more practical. Learning is a top use case for our voice model. People love to use it to dive deep into topics, get quizzed before exams, or learn new languages. For me, I've been learning Korean recently, and I found that speaking it and understanding it is incredibly challenging. But let's see how voice can help.
To do that, I will enable this study and learn mode, a new mode feature we're launching in ChatGPT that guides you step by step towards deeper understanding of a subject. Now, we can talk to voice. Hey Chad, I'm learning Korean. Could you help me practicing it? Let's say 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 I'm a beginner and want to hear every word clearly?
Now, let's level things up. Could you speak it up 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 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 4o demo.
Yes. Great. So today we would also like to announce a new feature and a set of features that make ChatGPT 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 ChatGPT such that it's more supportive, or it's more professional and concise, or maybe even a little bit sarcastic. And this lets you interact with ChatGPT in a way that's consistent with your own communication style.
But the way that ChatGBT sounds and the way that it looks is just one part of making ChatGBT 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 ChatGBT 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 ChatGBT getting to know you more and more over time. And this is our aspiration for ChatGBT to understand what's meaningful to you so it can help you achieve your goals in life. ChatGBT has already been so helpful for me. I'm training for a marathon right now and ChatGBT is helping me pull together a personalized running schedule.
But ChatGBT 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 ChatGBT 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 ChatGBT access to my Gmail and Google Calendar. So, it just works and it's easy here. But if you hadn't, ChatGBT would be asking you to connect right now.
Let's see what ChatGBT is doing. Okay, that was pretty quick. So ChatGBT has pulled in my schedule tomorrow and without even asking, ChatGBT 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. ChatGBT 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 redeye 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, ChatGBT is getting more useful and more personal. We're really excited for you to try this out next week.
Thank you so much for that. Great. So, we've seen a little bit about features that we've enhanced. Here to talk a little bit about the research that went into ChatGBT and the safety that made it more deployable, we have Sachi and Sebastian.
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 o3 and o4-mini.
We've also completely overhauled how we do safety training. So our old models, the models would look at 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 display or they could be trying to cause harm with this kind of information.
And so for this kind of prompt, o3 over-rotates on intent. As you can see, this particular prompt is stated in a way that's relatively neutral, has a lot of technical details. So we can see that o3 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, o3 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 continue the conversation in a more safe way.
So let's look at that same technical prompt that o3 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' responses. 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 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 high-quality 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. It 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 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 deepen.
In the future, AI systems 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 ChatGBT 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 Carolina to share their healthcare journey. Thank you so much for joining us.
So to start off with, could you tell us about the journey, the healthcare journey that you've been on?
Yeah. 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. 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 screenshot of the report and put it into ChatGBT 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 ChatGBT throughout?
I've used it in so many different aspects of my journey, but one of the ways that I found it most 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 ChatGBT to gain knowledge and understand the nuances of my case. And again, within minutes, it gave me a 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 30 minute 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 ChatGBT. In this moment it'd 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 Carolina's health 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 the promise of AI in healthcare isn't in just 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 you think?
I've been so mind-blown about GPT5 and its capabilities. 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?
Just think long enough, but it is very thorough. And more importantly, it feels more like a thought partner 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 GPT5 and GPT4 had done a great job. It had translated, explained what these words meant and helped in a way we can understand. But GPT5 seemed to understand more of the context and the question behind the question like why would we be asking biopsy results? And so it 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 we really started to pull together a complete 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 Carolina 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...
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 eight months ago.
I think we're incredibly excited for that, too. Thank you so much for coming to share your story. We're pleased that CHAP has been able to be helpful to you and we hope that the new version will really be able to help a lot of people. We wish you the very best.
Thank you. And I'd like to hand it over to our president, Greg Brockman.
Software engineering is already fundamentally changing and GPT-5 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 it 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, that you can talk to them and they actually do what you want. They can really amplify what you're able to accomplish and 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 GPT-4.1 and o3, but GPT-5 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. It's incredible at front end. It makes very beautiful visualizations and interactive games and you know, you've seen some of this in the live stream so far and you'll see some more upcoming. But it's just really amazing to see whatever you imagine coming to life.
It's extremely good at instruction following, very detailed instructions, being able to accomplish, you know, sort of when you have something very vaguely specified, inferring your intent, or something very detailed specified, actually following it. And it's also 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. But so we're making it available not just to developers to use to write their own code but to build novel applications. So we're putting it into the API, and to talk about that is Michelle.
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: GPT-5, GPT-5 Mini, and GPT-5 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 between a bunch of models and you can use GPT-5 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. 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 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 expression or even a context-free 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 preambles, and this is the model's ability to output an explanation of what it's about to do before it calls tools. This is not super new, but o3 didn't have this capability and in GPT-5 it's supercharged with extreme steerability. The model is able to follow instructions about these preambles 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 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 terse or expansive the model is with its outputs.
GPT-5 is a state-of-the-art coding model. On SWE-Bench, a measure of Python coding ability, GPT-5 sets a new high of 74.9%, versus the 69.1% from o3. On Aider Polyglot, which is a benchmark that covers all sorts of programming languages and not just Python, GPT-5 scores 88%, a stark improvement over o3. You've also seen that it's incredible at front-end web development. And so we've asked human trainers to look at outputs from GPT-5 and o3 and pick which they prefer. And they prefer GPT-5 70% of the time for its improved aesthetic abilities, but also better capabilities overall.
But GPT-5 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 TA² 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, this case in the telecom industry. So trying to solve 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 GPT-5 scores 97%.
GPT-5 is also state-of-the-art on general purpose instruction following. It scores 99% on Collye, which signals a great departure for this benchmark for us. It also scores 70% on Scale's multi-challenge benchmark, up 10 points from o3. This is a measure of multi-turn instruction following. Finally, the instruction following eval I actually prefer the most is one we've built in-house. It's based on real API use cases and for that reason it's a really good measure of how GPT-5 will perform in your application.
On the hard subset of this, GPT-5 scores 64%, up from 47% from o3. A pretty meaningful improvement. So we think it'll perform quite well in your applications.
We're also bringing GPT-5 to a longer context window in the API. It's now got 400K of total context, up from 200K from o3. But it's not enough to just release a longer context window. We wanted to make it more effective and usable. And GPT-5 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 Graphs Walks BFS Metadata, 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 in this model. We're also open sourcing a new long context eval called BrowseComp 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 GPT-5 is the best model for developers. It was trained with a focus on real world utility and less so on benchmarks, but we happened 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.
Thank you, Michelle. As Michelle was saying, the benchmarks, they're exciting numbers, but we're starting to saturate them. Like when you're moving between 98 and 99% of 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 and 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'd like to welcome Addie and Brian to the stage.
Thanks, Greg. I'm Brian, a solutions architect on the startups team.
I'm Addie, a researcher on the post-training team.
To recreate the ideal pair programmer, you need a model that understands best software engineering practices, but has a personality that just feels right to work with. For GPT-5, 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 GPT-5 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. And afterwards I tried to have GPT-4 o3 fix it for me and it couldn't. So while we were testing GPT-5 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.
All right, let's hope for better luck than with o3.
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 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. 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 in Holland?
You did. I couldn't do it.
I mean, I was busy. So, okay. So, continuing on, it's starting to figure out where it's going to sort of figure this out. So while this is going, let me tell you a little bit about how we trained GPT-5 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, it's been really amazing to see the team really doing the grind of going and seeing how this model behaves in practice, figuring out what people really want and putting that back into model training. That's something that I think has been a real focus for this model.
It's been pretty great. So while this is fixing, the other thing that we did 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 set it off, came back like 45 minutes later, it just finished and we tested it out and it ran the first time. It was pretty surprising.
That is magical.
Okay, so it made the edits. It looks like yeah, it found the right problem. And right now it's actually, it's running lints, but these lints are actually not related to this bug. So, it's going to ignore them. And then it's going to run a build. It'll run tests if there are any. 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, GPT-5 is totally tunable. You can steer it with system prompts or Cursor rules. You can change its verbosity levels or reasoning levels to match your tasks. And if you get stuck, ask it. GPT-5 is actually really good at modifying its own prompts by metaprompting.
So after using this for the past few weeks, it really feels like we've achieved state-of-the-art zero-shot 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 GPT-5 has come when it comes to coding personality and steerability. I'm really excited to show how great GPT-5 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. I have some data that I'd like to visualize about the company. And I'm going to ask the model to make me a dashboard.
So, you'll see here that I'm being specific about the audience. So, the target audience is the CFO. I said, create a finance dashboard for my startup. And I've asked it to be beautiful, tastefully designed with some interactivity, and to have a clear hierarchy for easy 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 create-next-app to make a Next.js project.
