OpenAI's Thibault Sottiaux on GPT-5.6: multi-agent Ultra mode, 3x faster computer use, and a 'Jarvis-style' voice experience

Speaker 2: lot. Cheers.

Speaker 1: Let me tell you about Shopify. Shopify is a commerce platform that grows with your business. Let's let's you sell in seconds online, in store, on mobile, on social, on marketplaces, and now with AI agents. Jordy, there is a story. Oh, wait. We actually have our next guest.

Speaker 2: Guest of honor. Now,

Speaker 1: Thibault from OpenAI. He's the head of core products and platform. Thibault, how are you doing? Congratulations on the launch.

Speaker 5: Hey. Thanks. Doing well.

Speaker 1: Give us the highlights

Speaker 3: I did

Speaker 2: sleep last night. Do you sleep do you sleep at all?

Speaker 5: Yeah. I do sleep. Currently, have, like, you know, maybe five to 10 war rooms going. So Okay. You know, it's just like a little bit it's intense. It's a sport.

Speaker 1: Yep. You know, but we're

Speaker 2: and you guys like to make it hard on yourselves by launching every every time there's a launch day. It's like, you know, 15 new things. So it makes sense that there's, you know, close to equal amount of war rooms.

Speaker 1: Yeah. Let's start with 5.6 Soul though. I want to I want to I want you to identify for me like what is sticking out? What are the most cutting edge capabilities that really stuck out to you? The latest unlocks from the frontier of the actual model, then we can go into codex and voice model and everything else and how things come together and how these are used. But first, just from the raw model capability, what was most impressive to you? What was most exciting?

Speaker 5: Yeah. It's actually really hard to answer that question, Chris, because when we were looking at the benchmarks, trying it, we were just blown away by it all. It's just better at coding, better at cyber, better at everything like long context, producing documents, better at having taste, website generation. And then for the first time, we also really cracked, I think, multi agent setups, which we shipped as the ultra mode. And when you just see that going and, you know, you've got, like, eight agents collaborating together, communicating, and, you know, getting the same work done, like, faster. It's just you just feel like, wow. Know, this is, like, another way to scale test time compute. Yep. But overall, just amazing workhorse. Feels way, way better than Five five and anything else we've produced so far.

Speaker 1: So in January, there was a project that was getting some attention called Gastown talking about all these different sub agents. You had poll cats and all sorts of different abstractions. It felt highly technical and it seems like Soul Ultra is a way to abstract that away. Is that deliberate? Is there I guess the question more broadly is like, is the level of prompt engineering, are we leaving that era? Or will there always be some cycle of if you get really good at using SOL Ultra, you'll have a better experience because you'll be able to give more fine tuned, fine grained prompt and direction to the model.

Speaker 5: So one thing that you see as well with SOL is its uncanny ability at understanding human intent. Mhmm. And, you know, you need shorter prompts. You don't need to explain yourself in that much detail and sort of, like, you know, just gets it and then goes and does, like, a very complex thing. Yeah. You know, you saw the prompt for, like, post training the Luna model, which is, like, super crisp, and then it does that, like but it actually worked for many days. Yeah. And this is also, like, with us launching Chat 50 Work. It's it's about making it accessible for everyone and, you know, you don't need to have, a PhD to use this model. It's just like it should just behave like just like another super, super smart human and, you know, just kinda get you in the moment, and that's what we're striving for. Of course, if, you know, you really push it to the limit, you're always going to find new setups, and this is, like, also a very exciting space. You know, we continue to develop, like, also codecs in the open source, and then we're seeing, like, you know, all sorts of sorts of novel ways to set up these agents and models so that you can get results in cybersecurity and all these other more nuanced and complex things. But for everyone, you should just feel a ton of power out of the box.

Speaker 3: Mhmm.

Speaker 2: Talk about what was important at the product layer. Fundamentally, what I think people want out of products is to just be able to talk to their computer like a really smart coworker and be able to get things done. But then you're dealing with, you know, so many users over here, millions of users over here trying to combine it and condense it into something that's simple and obviously simple things end up being, you know, exceptionally complex to actually create.

Speaker 5: Yeah. If you look at it, it's deceptively simple. You can open it on your phone. It's ChatGrniche work. You just toggle it and then there you go. You connect it to the things that you already have, your email, calendar, your docs. And then suddenly you're like, okay, wait. I can ask it to process all of this information that I had over there that I had to manually do all these things myself and it can just do all of that. And it's just like on the go and it's like on your phone in the Chateapp tab already have installed. I mean, that's the beauty of keeping it very simple. At the end of the day, we want it to be just a normal conversation between you and the agent. This is also why we decided to ship it just in chat with you.

Speaker 1: Mhmm. Talking about progress in computer use. What is actually driving progress there? Is this just something that sort of comes for free with scale and model advances or is there deliberate data collection that's happening and some sort of flywheel that's unlocking new capabilities in computer use?

Speaker 5: Yeah. We've done a lot of effort, bespoke effort on Windows, Mac, and, like, mobile computer use, also, like, phone use as well. And so there's an entire team working on this. It doesn't just come for free, but what does come for free is, like, every time we push the efficiency frontier and the model gets, like, you know, more efficient, like, at thinking and acting and it just, you know, costs less tokens and it gets compressed in time, it also gets better at computer because it reduces the latency and reduces the cost. And so the two compounds, like, you know, we have a lot of gains that we're getting from, you know, also like visual understanding. Every time, you know, it improves, it's like the model just gets more precise so it doesn't have to correct itself. And it was like, maybe it misclosed the button. They're just like, oh, it was like, wait, I have to redo that. So every time it's, like, more accurate and, you know, more token efficient, computer is definitely benefits from it. And when you compare it to five five, it's, like, you know, it's just really, like, three times faster. So, you know, we're not at all hitting a wall here in, like, how fast we can do computers.

Speaker 1: Yeah. Can you talk about how the role of member of technical staff is evolving? Because you're you're you're talking about Soul Ultra going off and working for days at a time. And at a certain point, your job sort of evolves to if you have a launch tomorrow, don't kick off a task that takes four days. Even if the model is capable of it and will deliver something great in four days, you need it tomorrow. And so you have to size your workloads appropriately. How is the how are you thinking about sizing work and and actually delegating the right the right chunk of work at this stage?

Speaker 5: Yeah. I find your question very interesting because it actually highlights a shift in our thinking over, you know, since we had five, sixes, you don't really instruct it necessarily for a task that's going to take four days. You tell it all the information that you have. So, you know, you're like, Hey, I have a launch tomorrow. Yeah. Right? And then, keep track of the time and, like, understand that, you know, the PR needs to land. Like, you know, I'm in that or, like, by 2AM. It will just, like, reason over it. Yep. And you're not the one that needs to manage, like, know, all of that extraneous, like, complexity. Yeah. And so that's it. What you're seeing as well is, like, you know, your relationship with the agent, like, changes over time as it gets more intelligent. And you're just like, oh, yeah. I can just talk to you, like, you know, another, like, you know, super start super smart human.

Speaker 2: Yeah. Yeah. Talk about the efficiency of the model, what work went into that, why why it matters, you know, what kind of conversations you're having with, you know, big customers, all that stuff.

Speaker 5: Yeah. So what matters a lot right now is sitting at the frontier and, you know, getting the max capability when you want it, but also for your normal average day to day task because, you know, being super efficient. And not just for latency, it's just because also we're seeing and so, like, we had this era of token maxing and then, you know, we've been talking a lot with, you know, old old companies and enterprises that we're working super closely with, and then they were like, oh, it's just a little bit, you know, maybe out of control. It's like, you know, what we want is, like, you know, were in a highly efficient model that is, you know, steerable, controllable. We want the the right, you know, spend control, dashboards, and so we also have all of that. You can you can look at your spend, understand the ROI, but also you can rest knowing that, you know, this is actually like a super, super efficient model. And so you get the job job done with, you know, way way way fewer tokens, which, you know, to you, means that you have to pay less for the same results, which is super important. This is really the theme, I think, of the year is being on the efficiency of performance and cost.

Speaker 1: What about if you want to spend more for faster performance? What does the future of either ASIC enabled, Cerebras enabled, Spark and fast mode, what do you want to see develop there either immediately or over the next couple

Speaker 5: Yeah. I think what we are truly working towards, like, it's a buffet of options. Right? So Mhmm. For for your normal, like, you know, interactive task is, like, you know, you're going to use five, six sole, like, on medium or on high, and you're going to have, like, an amazing time. If you have a really hard problem and you're trying to, for example, find a cyber vulnerability in something, and so you're gonna run ultra and you're gonna run it, like, for two days, and it's just gonna, like, leave no stone unturned and, like, you know, invent novel techniques and, you know, you're going to be, like, absolutely blown away by what it comes up with. And but a lot of times, you also just need speed, you know, like, for some of the subs that we're dealing with is, like, you know, we love working off of, like, the Cerebras version of this, which is, like, at, you know, about, like, you know, 750 tokens per second, which is an order of magnitude faster than the the default version that we have on the on the API and in the product today. And this is just really situational or if you just want the very best. Mhmm. There's absolutely no compromise. It does come at a cost.

Speaker 1: Of course. A lot of

Speaker 2: people were feeling left out this week that weren't in the early access program. What makes a good early access partner? I'm sure your DMs are just felt like, you know, people that want access to the next set of models. But what makes a good partner to the to the to the product and the research team?

Speaker 5: Yeah. We really try to go as broad as possible. It is quite a bit of effort to manage and then also, like, we're getting all that feedback and incorporating it and we work very closely with the the the folks in early access. For us, it's just really about realizing whether it is as good as we think it is. Right? You know, you're, like, so close to the model, you train it, you know, you've incorporated, like, all that feedback, all your dreams, visions into this model, and then you've played with it for a little bit. And then when we give access to, you know, folks outside of OpenAI, it's, you know, the first time where we have, like, an unbiased look, you know, where people are using all sorts of models and different harnesses every day. So it's just kind of like this awesome, you know, is it actually as good as we think it is? It's like, you know, what are the things that we missed? And so it's that, you know, high high bandwidth engagement, good feedback, and then, you know, that sort of people who have shown to be unbiased in the past and talked honestly about all sorts of models and all sorts of harnesses.

Speaker 1: How are you thinking about the tradeoff between mobile, cloud, desktop, the Mac mini that went mega viral last year? Do you think that we'll stay in a hybrid pattern for the foreseeable future? Is it a person by person basis? Do you have a grand unifying theory of how agentic work happens in the future?

Speaker 5: Yeah. The way we think about it is no compromise. Mhmm. So you want to be able to use the same same AI partner, you know, like on your phone Yeah. On you know, like on the go. It's just like, you know, I go walk in the park. I want the exact same thing.

Speaker 6: Mhmm.

Speaker 5: I want it on my laptop. I want it, you know, at home maybe like running in a Mac Mini. Is more that it needs to be able to have access to all the things that are important in my life and, you know, not be constrained by the physical, you know, boundaries of like, you know, it's just like, hey, I started this prompt or I started this conversation on my phone and it's just like now it's stuck on my phone. Yeah. But, you know, we want it to be uncompromising and so your AI your ideal AI partner, I think, just, you know, has access to everything all the time and just, you know, processes the information as needed and then, you know, can act in a safe way and controlled way so that, you know, you always understand, like, what it's trying to do. You know, if there is, like, something risky, you can, you know, approve it or, you know, ask it to change tack. And there, I also think, you know, like, the mobile has, a big role to play. Right? So if you're if, you know, five, six hours, you know, busy, like, you know, working on something and then, you know, you just go out to dinner, it's like, you know, it should ask you for permission to do something

Speaker 1: Sure.

Speaker 5: When you're there. And, you know, you don't don't need to be stuck on your laptop.

Speaker 2: Yeah. Fast forwarding six or twelve months, how is how how important is voice to someone's day to day experience with ChatGPT work slash codecs?

Speaker 5: I don't think you would need to fast track it. You know, we shipped yesterday.

Speaker 2: No. Know. But but, you know, you assume, like, you know, oftentimes, like, something ships, you know, and it takes a little

Speaker 1: Everyone has to go on holiday break.

Speaker 2: Yeah. Like, you have three day weekend.

Speaker 1: Three day weekend. Then everyone can test out the latest and and and integrate it into their workflows.

Speaker 5: Yeah. You know, open a chat with the app, like, using it, latest voice. We also demoed it in

Speaker 3: in the

Speaker 5: livestream this morning, and it's very sort of super enjoyable, like, magical experience when you first experience it, but also, like, know, on the fifth time as well. It's going to be part of, like, know, day to day experience, you know, of, like, how you work with these systems. We don't we don't have it yet in the desktop app, but this is something that we're working towards. And when when you experience it, is it it's a it's like a modern day, like Jarvis. Right? It's like, you know, you just talk to it, you just walk in your room, and, you know, suddenly it's just like doing things on your computer, you know, with the same level of, like, precision and power, you know, that you currently have over text.

Speaker 1: Yeah. It's amazing. Fantastic. Congratulations. Thank you so Hopefully, you can get some sleep. I'm sure Many

Speaker 2: many big days

Speaker 1: to come. Back to the one of five war rooms, whichever one you'll go to next. Have a great rest of your day. Congratulations. And we'll talk to you soon, Thibault. Cheers. Have a great one. Goodbye. Let me tell you about Railway. Railway is the all in one intelligent cloud provider.