【特別編】OpenAI DevDay 2026 ライブ視聴会：発表のすべてにリアルタイムで反応

All right. Hello, hello, hello, and welcome to our little Deb Day watch party here at the Neuron Live. We're just getting everything set up, get folks a few minutes to straggle in. We're getting our views set up here, too, so we can watch. We've debated on, over the past, the best way to do these. I mean, like, we can co-stream it here, or we can all just actually watch it and talk amongst ourselves here, whichever way everybody thinks will be, like, the best. We should at least have a version of it up that is muted. Yeah. And not subtitles if we – Well, I don't want to – my fear with that is it's also someone else's IP. It's true. It's true. But I can't hear you at all. Like you're there, but it's almost a –

I'm going to check your audio settings. I'm dimed. Well, shout out in the chat if you can hear me. Can everyone hear me? Now I can hear you. Okay, I just have to talk louder. You sound normal now. You sound normal to me at max volume now. Got it. Okay. That's strange. Grant, does your mic have a knob? It has a knob, but it's for headphone volume. Oh, okay. Yeah. Cool, cool. I mean, I can bring it closer to my face. Ooh. What up, Dadrock? As folks are tuning in, let us know where you're streaming from and if you are hyped for this dev day. Yeah. Dadrock, where are you writing in from? Your watch history is paused. Why is my watch history paused on YouTube?

That makes no sense. All right. Where is OpenAI? Dev Day Keynote. All right. I want to also... Fuzzy, fuzzy, fuzzy little dudes. We got Muse. We got Blocks, little fun agents. We've got what looks to be coming from here. Manus dropped another agent tool yesterday that was also kind of fun. Oh, yeah. Yeah, we wrote about that in the newsletter this morning. Okay, we got Bellingham, north of Seattle. Nice. I think I've been to Bellingham. Or I have a friend from there. Raleigh. Nice. My sister graduated from Duke University in Durham, so went out there for her college graduation and got to check out Raleigh. It was pretty sick.

All right, looks like we got 22 in the chat. Let's go. Let's get those numbers up. Tell your friends. Oh, so Grant. Yes. What do you expect today? I am expecting dot, dot, dot. Yeah. There is supposedly, allegedly, this is not a leak, I'm only speaking to what I've seen X.com rumors about, that there is something called DOT that is supposed to be OpenAI's personal agent tool. Yeah, and if the rumor mill is to be trusted, and it rarely is. It rarely? I feel like it's pretty spot on most of the time, actually. I say that. It does pretty good. It's more on top of it than I would think. Let's put it that way. I think, yeah, I think we'll see Dot, obviously, because there's dots floating all over my screen. And I'm really excited about that.

Oh, nice. That's awesome. And TikTok. I think this will be cool. I also think there will be a new cyber model today, it sounds like. I saw that mentioned the other day I'm not sure if we'll get a new model I read online a few minutes ago I was looking at an article about because you know they made these announcements that got everybody up in arms around the $200 plan but the rumor is that they're going to be cutting API credits per dollar in half so the idea being to motivate people to use the api connections because as it is the truth is the 200 plan is an insane deal like like 20x is a lot uh when you think of over the 20 plan like 100 plan is 5x and it's four times that for 200 bucks which is amazing for subscribers but is kind

of a stab in the back to developers i would say like if you're a developer you're being kind of Hang on, I've got to turn the volume down on my YouTube. You're not hearing any volume from this, right? No, no. I couldn't hear you, though, because mine was still coming out of my speaker. I think I was just going, mm-hmm. So you didn't miss much. I'm nodding along. But, yeah, so I'm curious to see how they structure that. But, like, the truth is devs have gotten kind of the short end of the stick when it comes to how far your dollar can go when buying intelligence for a project. And they're also kind of the lion's share of the revenue as a general rule. Happy Dev Day. What up, Blind Dragon? Happy Dev Day. We know you. We've seen you before. Yeah, totally. I think that's, like, the controversy leading into this. We'll see how they handle it. Do you think they handled it the right way by leaking this the night before and then having everyone mad and then revising the statement leading up to it?

I think so. I think you let people get angry ahead of time and then hope to come out with something that makes it good tomorrow. I think it's better to end on a good note. Right. And we'll see. Supposedly they've got more than 20 announcements coming today, which is so much. I'm anxious to see what they are and if they're substantial or if they're small or what the case is. You know, but we shall see. Yeah. Let's see. What else do I expect? I expect something around how their workspace product works. Yeah, I think that's possible. I think there's already some changes in codecs this morning. Like I noticed there's now a developer – oh, what is it called? One second. Open AI Developers is an option on the left side. First off, they added that bar down the left last week, and I love it so, so much. I love getting away – the business account doesn't have the bar down the left on web yet, but my personal ChatGPT account does, and I love it.

Grant looks like, I've never seen said bar. What are you talking about? Yeah, I'm like, what do you mean? You know how cluttered the left side of your screen is in ChatGPT? Yes, it's kind of a... It's now an icon bar over down the left side of the screen where you can just open what you want. Interesting, interesting. Way cleaner, way cleaner. But it has this opening iDevelopers thing where it looks like they've worked in basically all of the parts of your API platform, much of it at least, directly into Codex. So, like, you can create an API key right there in Codex. You can build an agent right there in Codex. You can explore things to build. You can optimize for better prompt caching. You can upgrade models, optimize your API implementation, build a voice agent with GPT Live, all right inside Codex now, which is really pretty baller. Yeah. And a customized function. They brought plugins and skills into a cleaner look.

Nice, nice. I'm not sure what else. Let's go back over here and open a new chat. We got a countdown on the screen, Corey. You see this? Oh, now it's a party. We got a countdown, y'all. I'm seeing all the models look the same still. Usually the models don't drop until after the speech anyways, if there are models. I mean, I don't know. I think there will be a cybersecurity model. I don't really know what else to expect on that front. Like last night, the Wall Street Journal reported that they're pausing Astro 6.1 for cybersecurity concerns, said it. I think the quote was didn't quite meet their safety bar, something to that effect. Yep. Yeah. So we'll see what the story is there. yeah we're probably not going to see a new model unless they unless that's some subterfuge yeah yeah some uh smoke and mirrors to try and throw people off but i think after a story like that runs and then you release it then you would that probably would look pretty bad on you i would

think yeah i think so too i mean that's what they want it doesn't mean there won't be something else They keep floating a lot of O's around, and I don't know if those were to imply dots or if those O's were maybe a new level of O series. Yeah, I think Omni. I think that's what that implies to me anyway. It's some sort of Omni. Yeah, which for folks who don't know, an Omni model is a multimodal model, so it can do text. It can do potentially, I don't know about voice, but video and images. It can kind of do all sorts of those things. Oh, you're thinking about a 6-0. I was thinking an 06. More like a 1, 03. Yeah. Yeah, I was thinking more an Omni model, but I mean, I don't know. I'm just guessing. Yeah, I am too. I'm absolutely just spitballing here. It could be anything. It could be nothing. I think they kind of collapsed the 06, like those ones, into just like now all models work that way. Into what is now all models are that? Yeah, yeah.

And for folks who don't know, that's like the 01, 03 series. Those were the first models that had this chain of thought reasoning that you could see in there. I guess you could consider them large reasoning models. So LRMs, although you only really see highly technical people use that term, at least in my experience. Yeah. Yeah, I think so, too. Looks like we're getting down 35. Wow, that's confusing. Either way, I'm anxious to see what the new tool is. I'm also going to be watching the App Store because I assume we're going to be looking at maybe an update, maybe an all-new app altogether. I also heard tell that we should expect a full Codex remote experience. Oh, really? Which would be pretty baller. Yeah. Yeah, because it's kind of a hacky workaround right now. All right, seven, six, five, four, three, two, one. All right. Okay. Okay.

Should I turn volume on? Let me know if it's too loud. I'll do. I'm listening on my phone. I've got my computer down. Tell me if this is okay. Can you speak? Yeah. Can you still hear me over what's being said? Yeah. Good morning and welcome to Dev Day. It is great to have everyone here and everyone joining us online. Thank you for coming. I think this is going to be our best Dev Day yet. We have shipped a lot since last Dev Day and you all have built amazing things with it. You find uses for these models that we hadn't thought of You've also helped make all of our products better. You have sent us thousands of requests, mostly on Twitter with ideas for what we can improve We try to read all of them and we work through them as fast as we can You asked us to bring Codex to Linux So we did that You asked? Yeah A lot of people like that one You asked us to work across multiple folders and one codex projects, and now you can do that, too.

And you asked for codex on your phone, so we brought it there. Boy, that was a changer. It gives us a lot of joy to build things that you find really helpful. And the next thing is one of the things that you've all asked for the most. It's something we've been quite excited about, too, so let's get into it. We're kicking off today with a product that we believe is going to feel like a whole new way to work with AI. this is the real deal version of ai that we've always imagined for ourselves it's like an ai helper that always has your back inspired by the cool versions of what we all watched in movies growing up it can hugely magnify your ability and expand your potential working 24 7 to take on things that you don't want to do so you can focus on higher level work now you may have seen early signs of what this could look like in virtual assistants that help you book a reservation or buy a plane ticket. And these things are useful to hand off, but they're distributed relative to what the technology can do. And so we want it to be way more ambitious.

Think about that AI from the future that you imagined. It knows what's going on. It knows what matters to you. You don't have to tell it every little thing to you. For sure, Crock-Clock competitive. We want AI to feel like that. We want it to feel like magical things that it just does for you without you having to ask. So today, we're going to be Just crackpot and a user book. Dot. Dot. That's sexy. I like it. Dots are remarkably capable, always-on agents, built to handle really anything you can do. Like a little beret, the sunglasses. They bring AI into a whole new form factor to work and give you more time and attention so you can do the things that you care about. Let's eat dots. Turn up the box for this part. Hey, I'm your dot. I'm new in ChatGPT.

I'll be here to handle everything that comes your way. Ready to jump in when you need me. I'm going to call you... I'm going to call you Alfred. You got it. I'm Alfred. I'll keep things moving and check in when something important needs your approval. Love it. Let's get to work. I've been getting your website ready for the full launch using your approved design. Okay, perfect. The latest consumer experience numbers came in on Teams. I've been updating your board meeting deck. Sounds good. So, I cut the app migration in your Google Meet Notes. I've already mapped out the changes. Get it started, but make sure you first check every flow against the current product before I review. Also, update on wedding planning. Bad news, the cake vendor canceled. Good news. Not the cake vendor. Lifesaver, when can we do a tasting? Saturday at 11 a.m. works for both of you. Pro tip, don't get a cake at your wedding. Do ice cream. Hey there.

Your website's updated with the new designs for the four drop. Check it out. Nice. I like that. Just swap the first and third photo, and then you can publish the sites. Done. You're live. Your board meeting deck is ready Product fees is climbing And revenue is up 51% year over year Okay, this looks good Except for change it so that the consumer Experience is first and then get into The growth numbers How's it looking? The rebuild is faster and testing With no regressions found so far The preview and results are in the PR Amazing, share the results with the team on Slack And get the stage rollout ready for my approval I'll flag anything that changes Also I book the kids after school lessons as soon as registration opens. I love that. Thank you. Tomorrow's finance review just got moved to the same time as your daughter's recital, so I'm going to see if we can reschedule. Okay, yes. Please. I can't miss that recital. I'll sign a letter. I'm not doing notes now. Good call.

I was just immediately thinking about the classic parents. I love the screen form track they were showing at you. I thought for a second there they were going to do a... What feels new about my dot is that it gets to know how I like to work over time and where I want to focus my attention. It can do things like review what came in overnight and ping me in the morning with anything urgent before I start my day. I've been really surprised at what a difference this makes in my life. I feel like I've finally gotten some of my attention back. I no longer feel quite as addicted to my phone in the same way. It's the way I've always wanted to work with AI. Of course, dots can take care of the small things, like those plane tickets or the time for dinner, or help with the home renovation that you've been putting off. But because Dots are powered by Astra and a lot of other things we built around them, you can delegate ambitious pieces of work the way you would to a high agency engineer or a chief of staff that you work with. You just give Dots a responsibility, like monitor incoming bug reports, tests and fixes and draft PRs, kick off next year's budget planning cycle,

and keep all the teams on track. Find out what's slowing down our app, build a fix, and show me how much faster it runs. And your Dots will just get to work and keep working. So here's a good example. for dots. Get our app off this legacy API before it shuts down. That is a difficult thing. The API might be used all over a code base, and it's not obvious what's going to break as you start making changes. But dots can trace those dependencies and figure out everything that needs to change. They can write the code, run the test, and bring our team pull requests to review. Think about how you would have done that the old way. How many people would have been involved? How much time would it have taken them? What else could they have done instead? Now, imagine having dots do that for everything you can think of. Dots on board with everything they need to get to work. They have their own computer in the cloud, their own browser, and they can write and test code as needed. You'll be able to start with one dot today, and in the future, if you'd like, you can work with a whole team of dots.

You also don't need to be in ChatGPT to talk to your dot. Coming soon, you'll be able to send it a text message. It'll text you back. It can text you updates, flag when something's changed, or ask you to weigh in on a decision. You can even give your dot a call and talk with it directly. Dots also work with more than 4,000 apps across ChatGPT. When you onboard your dot, it gets to use the same plugins you've already connected. It can just follow along with your work wherever it happens. Through all of this, of course, you stay in control. Dots are part of my app, our most aligned model, and they have built-in protections and safeguards. You can set the boundaries on apps and computer use, and you can customize instructions for any actions your dot takes. So this allows you to trust your dot with as much responsibility as you're comfortable with. Okay, so that's dots. Dots are very deeply tied to you, your context and your workflows, and your desire to get things done. But you also need a space to collaborate with your team and their dots. And the problem here is that almost all productivity software

was not built for AI and people to work well together. Facts. So we're changing that too. Next up, we're introducing ChatGVT Space. ChatGVT Space. This is a place where you, your teammates, and your agents can all work together. Space starts with a page like the one here you see on the screen. You can write a plan, do research, create images, generate data, or build something new. Dots can work alongside you directly in the page. You can bring in a dot and ask it to track dependencies just like you would with a senior engineer. You can tag a dot in a comment, and it'll jump in and get to work. Your pages and files all live together in a space, like they would in a drive. But spaces feel alive. You can give the page specific instructions, like check the AGI platform's Slack channel, and update it daily with findings here. Coming soon, we're introducing a new way to build presentations right inside Space 2, so you'll be able to share a presentation with your team, collaborate together with your dot and their dots right inside ChatGPT.

This will be a version of slides that could be easily read and modified by agents. So at this point, I'd like to welcome Holly on stage to show you a demo of how this all works together. Thank you, Holly. That feels like almost an AI first drive solution. Hi, everyone. I'm Holly, and I'm on the product team here at OpenAI. So Dots have changed how we work and even how we built the product itself. So today I'm going to show you how teams at OpenAI use Dots via a fictional music app we're calling Blossom. So this is Blossom Music. It's a fictional, cool app for creating playlists. And this is my Dot, Dottie. Dottie. So Dots have all of your connectors. You can add them to Slack. You can orchestrate your codex. And so Dottie is truly indispensable to me. It's a second set of eyes across everything I give it access to. And more importantly, a second set of hands to actually take work off my plate. So let's say we're gearing up to launch Blossom Music.

It's tomorrow, okay? Things are chaotic. But Dottie has my back. This morning, it proactively caught a bunch of things for me to look at. Our launch review moved earlier. It already updated my calendar. We shipped some new features to test users last night, and it's already looked at the feedback. And it looks like design landed a last-minute rev on our home screen. And Dottie's not only spotted this, but actually sent me the design to take a look at right here. So this is all context. I'd normally have to piece together myself or worse, things I would have missed entirely. But Dottie has it. So let's take a look at this new design. Okay, I like it. But this close to launch, I really want to be able to play with it, right? Because Dottie has access to this laptop and its dev tools, I'm just going to say, go build it. And no pressure, Dottie. We'll check back on this later. So here's my personal favorite part of working with Dot. Voice. My teammates make fun of me because I'm always holed up in a phone booth talking to Dottie.

But talking helps me think. I like to riff, brainstorm ideas. and because Dottie already knows who I am, what I'm working on, and why, I can get straight to the point. It's like talking to a second brain. So let's say I'm walking into the office, and I just need to catch up on last night's alpha. Are you calling Dottie? Good morning, Holly. Nice to hear from you. Hey, so I'm just walking in. Can you catch me up on just user testing from last night? Checking that now. I don't have to look like a weirdo in the grocery store anymore. Yeah. I'm just on the phone. Yeah, I'm just on the phone. Don't mind me. All right. How's it going? Still checking. This is classic. All right. I think maybe Dottie is having kind of a slow morning. I might ask it, too. Hey, I'm going to review with the go-to-market leads later.

Do you want to just drop an FAQ into the page? Sure, I'll take care of that. Okay, thank you. So, I love voice. I love jamming with Dottie one-on-one. But sometimes we know that the best ideas are developed together. One second. So, we've added a brand new place to collaborate with your team inside of ChatGPT called Space. Awesome. So when I hit slash here, you can see all of the different things that you can build. Interactive charts, sheets, live prototypes, anything you need to bring your ideas to life. And your team and your dots can actually join you in here. So let me show you what I mean. I can just tag in teammates like Renny or Rishab, but I can also tag in Dottie itself and say, hey, can we actually turn this into a bar chart, please?

And the visualizations in here are completely dynamic. I can filter by feedback types, and I can inspect these data points. And my dot can keep this up to date every single hour. Now, here's the FAQ that we just asked Dottie to drop in. And because Dottie has all my context, I can actually just, you know, hand it off things that would really slow me down. Like, hey, let's add in that stat about onboarding from that one engineer in my Slack DMs. And Dottie can find it. Or maybe something like, hey, let's incorporate more of Laura's feedback. Okay, so space is the place where dots and your team can jam together and stay in sync. But where Dot has really changed how we work is Slack. All of us are getting tagged all day. And sometimes an innocuous ask like reconciling launch specs can turn into a whole project.

Now I can forward the thread to Dottie. I can just delegate and say, hey, can you take this? And as you can see, my DMs are full of these little handoffs. I get to decide what needs my attention and what my dots can take care of. And by giving dots their own identity, they become extensions of our team. So here's something we totally didn't expect at OpenAI. People started treating their dots like they're delegates. And now we bring our dots into group chats from the get-go. So here, Gabe and Patrick are handing off tasks to their dots. When the dots come back with the results, everyone stays in the loop. And our engineers are taking this one step further. So here, someone's dropped in a session ID and a user bug into a feedback channel somewhere. But Gabe's dot picks it up, investigates it right away, and opens a PR with a fix. And this is a real example. Our engineers have their dots fixing dozens of bugs a day in this way.

It's been amazing to see how much dots were built by dots. Okay, so by now I've delegated a lot of different things to my dots, and it's freed up my time to brainstorm, spend time with the team. This is a live demo, and it looks like earlier there was a little bit of an error throwing up this codex thread. So I'm going to try it again. And what's happening here is that Dottie can actually use Codex on my laptop to build the app and launch it in the iPhone computer. Now, this might take a bit, but the amazing thing is we can start with the mock on my phone, run local execution, and get it to be a working app here. And then I can even ask Dottie to just run with it and ship the PR, all while I'm here with you. Oh, good call out, DadRock. DadRock says BBD 6.1 still is announced on the list. So dots are extensions of us. We can give them all of our context and memory,

and they can take work off of our plates. They can filter the noise, and when you add all of this together, dots really become extensions of you. It's not just about getting the same work done faster. Here at OpenAI, we've actually felt this firsthand. The dots are changing what we feel capable of taking on. And one of the scales to everyone, imagine what you can build together. So I can't wait to see what you do with this. Thank you. I have personally been waiting for this product for years, and we can't wait to hear the feedback from you all on how you're going to use it. Both dots and space will be available to all ChatGPT Pro, Business Premium, and Enterprise users today. and your dot is included in your plan and conversations don't eat into your usage. You can always work with your dot. It's available 24-7. One more thing for enterprise users.

We're also previewing specialist dots today. These are like virtual teammates set up by your company that will work alongside your whole team. They can help scale intensive work like accounting. marketing, legal, much more. You can give them goals and contexts, review what they produce, give them feedback to improve, and that will be shared across your whole company. We've been testing these inside of OpenAI, and they have been remarkably effective in many different roles. So for our enterprise customers, we'd love to work with you to get them set up and make sure they're really great. We're also working with Microsoft to integrate specialist dots into Agent 365. So this means businesses can manage their dots through the Microsoft tools that they already use. We believe that things like Dots, along with our models, will give people a whole new set of superpowers. You can get so much more done than you thought. Whether you're using AI to expand a small business, create more accessibility, or finding new ways to fight infection. We want people to discover that they can do things with AI they never thought possible. Yeah, this is key.

So that brings us to the second big part of today. We want to support you all with a powerful platform for building whatever you can dream of. There are three parts to this that we're really going to talk about. Models. We want you to have the best access to the best intelligence at the speed and the cost you need. Second, better tools. We want you to have the same tools we use inside of OpenAI, the freedom to build and run your products wherever it makes sense for you. And third, distribution. We want to help people discover what you build and connect you with new customers. Oh, distribution. That's cool. So first, let's start with the model. we launched Astra a few weeks ago it's the most capable and aligned model in the market but people keep asking us for two things make it cheaper and make it faster and we really didn't want to disappoint you we are launching GPT's 6.1 soul there it is you don't even know what it is yet we got to tell you why it's good first

this is a powerful model It has a great balance of cost and capability It gives you very near intelligence at a fifth of the price to help you code and build It extremely affordable to run In fact the cash input costs are a 95% discount on standard input. So this makes it much cheaper for agents to reuse context, iterate on solutions, and work through longer tasks. We think you will really love this model that is smarter than Astro in some ways, and it will become your daily workhorse. So we're very excited to get 6.1 Sol launched today. But sometimes you just want the highest intelligence at an absurdly fast speed. Rarely. So today we're launching ultra fast across the API, ChatGBT, and Codex. And we think you are really going to love this and not want to go back. Fast already gives you twice the speed at twice the price of standard.

But ultra-fast takes that to eight times the speed at six times the price of standard. An incredible 300 tokens per second. You know what? It's worth it. Let me show you, actually. So this is the same task side by side. We gave both models the same prompt. Create and launch a rocket with the dev day colors. On the left here, you have ultra-fast. On the right, you have regular. You can see how much faster this builds. The ultra-fast rocket is already taken off, and the regular rocket is not even close to being finished. There is nothing like this on the market today. You can use UltraFast now with Astra, and soon we're going to offer it for 6.1 Sol as well. So here's what these improvements mean for your Pro subscription. We're introducing Pro 500 with our highest usage limits and access to UltraFast in ChatGPT and Codex. Pro 500 will have 25 times the usage of Plus, and you can use it across partner apps with signing with ChatGPT, which we'll talk about later. We are also reopening the Pro 200 subscription today.

Pro 200 will continue to have access to all of our frontier models. And with GPT 6.1 Sol, you will have a model close to Astra quality that you can use a lot as your daily driver. Our goal is to give you the best combination of intelligence and capability at every price point in the whole market. One more exciting speed update. Today we're previewing our new Decisions API. The Decisions API lets the model respond in a fraction of a second. It's so fast. So it really feels almost instant to interact with. You can see here at using the computer and how quickly it's moving through these steps, this is not fed up. This works by giving our Luna model a predefined set of options to choose from, like routing a request, classifying an image, deciding what agency should do next, And by focusing the model on that choice, we can make it extremely fast while keeping capabilities like image understanding, broad language support, and safety protections.

We're very excited to see what you'll all build with this. So, thank you. I like it. So, OpenAI now offers leading models across intelligence, price, speed, voice, images, frontier models in every category. we want to be a one-stop shop where you can build, I'm really proud of our research team for getting the models to be this good, usable at this kind of scale, available at these price points, into the hands of developers everywhere. Now, there's a question of how the researchers can do all of this. Last year, around this time, Jakob and I made a somewhat bold prediction on a live stream, saying that we would see the first AI research intern in the next year. To say the world did not have a lot of confidence in that prediction at the time was an understatement. A few weeks ago, we announced that we hit that goal. We now have an AI research intern, a system that can take on a clearly defined research task

that would take a skilled researcher a lot of time and effort to complete. And we want to show you what that looks like and how capable these models have become. So for that, I'd like to bring up one of the researchers working on these models. Please welcome Tejal. Thanks, Sam. Hi, I'm Tagal. I'm a research lead for post-training, and I'm excited to share how these models are changing the way we do research here at OpenAI. Astro-unrelated models are great for math, coding, and computer use, and all of these matter a lot for research. Being able to use a great model for research helps us train even better and more aligned models in the future so we can bring capabilities that are more useful to you. For example, my collaborators and I work on computer use, and our models help make Astra a step change improvement for the industry in speed, performance, and safety for desktop and browser navigation. 15 examples, not sped up, of the model helping with legal tasks,

designing assets for a video game, making a dashboard, or even filling in your taxes. And our models help make this improved training possible. Those are real uses. One way this happens is our models actually help optimize the harness we use for computer use. So here, you have latency on the x-axis. We want latency to go down. and we have performance on the y-axis. And we actually had the models run in a continuous loop to find wins in the harness that would improve latency and performance. And then we pulled these into the real prod harness and post-trained on them. And this has resulted in already a more than 2x latency win that we've shipped to all of you. Wow. Even more importantly, and I'm very proud of this one, our models help make Astra much safer at computer use by improving our monitoring and our model behavior training to improve refusals in unsafe scenarios. And this has helped us achieve industry-leading performance on computer use stress tests. So our models are much less likely to make a mistake when navigating a desktop or browser and much more likely to be aligned with user intent,

which really matters for all of us. Beyond computer use, our models are accelerating our research program. Our team tracks token use within research, And you can see, especially after this summer, token use really skyrocketed internally. And we're seeing returns on this investment. For example, we tracked how often a model can carry out a task in research. And in January, the models were pretty successful on short, less than 15-minute tasks, but failed the vast majority of times on a task that took a day or longer. But as of July, and with that good, good, ultra-fast, our models have taken on over a third of day-long research tasks with no intervention required. So what does this mean? That's why they need the hundreds of billions of dollars. The mission of OpenAI is to ensure that AGI benefits all of humanity, and we want our models to help tackle some of the most complex challenges we are facing. Models like Astra have already helped solve over 100 math problems that were open for decades, and they're also helping accelerate creating new antibiotics for drug-resistant infections,

studying ancient languages and history, developing better renewable energy and improving energy efficiency, and accelerating industrial products in manufacturing. We believe our models will help make developers and users make very strong progress on some of the most important and ambitious problems for humanity. And we're excited to see what you discover and build next. Thank you, and back to you, Sam. We need not be afraid of our appreciation. Thanks, Rachel. That was awesome. We have a lot more stuff. So that brings us to the second part of supporting developers. Giving you the tools to build what comes next. we want you to have access to the same foundations that we have at OpenAI. This starts with the Codex harness. It powers Codex, work, and now DOTS. We have put a lot of effort into making this harness very efficient so that you get accurate results at the fastest speed with the fewest tokens possible. And it's open source, so you can look under the hood, thank you, and build it yourself. To power products like ChatGPT, we built this harness to run its scale for billions of users,

and that same scale is there to support your products too. Next up, we're making all of this available in the cloud. This is something I know you all have been waiting for. Codex is now fully in the cloud. Has it been up before? No. Yeah, it's not fully. Interesting. So this means you can start a task on your phone, pick it up on your browser or your desktop, right where you left off, no more lost threads. You don't have to worry about closing your laptop in the middle of a run. Your agents just keep working wherever you are. I have multiple computers. This cloud harness unlocks a lot of new possibilities. And one important example that we think you're going to love is our new Codex Security Cloud that we're launching today. We wanted to give defenders a better set of tools to harden their infrastructure. We see what's coming. And the new Codex Security Cloud runs in cloud environments. It continuously finds vulnerabilities and prepares verified fixes for you to review. We've added some highly requested new features like automatic deduplication, scheduled scans, a new interface. and you can start using it today.

Next up, we're launching a new Agents API to make all of this available to you. The Agents API includes great core components like the harness, hosting, memory, multi-agent controls, and much more. This is the same technology that powers codecs, dots, and all our agents, and it is now available in public beta, so you can start building. And we're bringing computer use to Agents API. This gives your agents a whole new set of powers. We want to show you a demo of what this looks like. So say you're building an agent that tests websites. This agent can now use a computer to open the browser, click through the pages, and test the flow. We also want to support you on your infrastructure of choice. It's been great to see how quickly our partnership with AWS has grown. We've worked together with them on Bedrock Managed Agents powered by OpenAI. you can build OpenAI agents in AWS right alongside your existing applications and data. AWS customers can access our frontier models and use Codex and ChatGPT work too.

Okay, here's a big one. When you build these agents, they need to work safely and protect your data. So today we're previewing OpenAI private intelligence to give you stronger controls over your data while using the most advanced capabilities of our models. What is private intelligence? GDR with private safety processing delivers frontier model safety without storing your content on our servers at all. And with private inference, we can even provide privacy at the time of inference. Together, we think these set a new standard for privacy and frontier AI. We've worked closely with some of our largest customers to design privacy and safety systems that meet their highest standards so they can apply this intelligence, and you can too, to their most sensitive work. yeah so does this work in theory all of this and everything you're building needs both speed and reliability like what we've been hard at work on this we've seen 100x growth in the responses api in last year 100x in the last year and we have still kept more than 99 reliability we've now cut the time to first token by 45 and we have set up tool calls and workflows by more

than 30 we will keep going on this but these improvements make this the most performant and reliable API available anywhere in the market today. So, thank you. That's big. I think it sounds like a thing you brought a lot of new things for developers to build with, and it wouldn't be that day without a demo from Ramon. So, let's bring him on stage to show all the new ways you can build with these new tools. Thanks, Don. I'm just taking a few photos of the waterfront. I'll be right there. Hey, Codex, can you help me? Oh, just the same buildings we're building. Sure, I can help. Tell me what you need. Can you welcome everyone and share a fun fact about Fort Mason? I'll be coming shortly from the left door, so get my demo ready. Thanks. Let me get that ready. Hello, Deb Day. Great to have you here at Fort Mason. And hello to everyone watching around the world. Fun fact, at San Francisco's 1915 World Fair, visitors could hear voices all the way from New York over the telephone.

Today, you've got an AI on the mic, a room full of amazing developers, and a few new tools to play with. Oh, there's Romain. All right, good morning, everyone. I love how we're building with Codex, even when we're not on our laptops, and kick off a task like this one from anywhere. Last year, we did not even have an app for Codex. Now, I don't have enough time today to show you half of the things we're shipping, so let's dive right in. it's been awesome to see how all of you have been creating 3d worlds with astra the past few weeks and i took a few basic photos of the venue to kind of recreate the same thing here so let's take a look what i really wanted to try is how we could change it together in real time with ultra fast so let's dive in The nudity is ultra fast for these demos, so it actually finishes. Can you second to start? Might I hear? That's so cool. Let's try again.

Live demos. It's not a live demo if something doesn't fail to work on the spot. There have been some great failures over the years. You see all my little dots down there? okay this is the one percent of down on the response this is the point one you know what i'll just uh tell simply tell it um fly me to the stage all right so uh what's interesting here is that i can interact with the the room with ultrafast so what i'll do instead here if we have little difficulties with the voice chat i'll go back here and let's do something else with this room now it's a little empty so i'll say uh why don't you uh fly uh the dots to their seats for instance would be nice because uh those are empty so far time to bring that and here they are so we're just changing this world in real time as we go pretty cool um maybe we can do something a bit more risky now uh and you know why don't we just

turn the camera to the room, to the audience, and I'll say, put the live stream on the big screen. And let's see what happened there. There we go. Wow. So if you want to wave at the audience, for those of you watching on YouTube, thank you for tuning in. Great to have you. So with UltraFast, you can build and really kind of change the world like this or build anything almost as fast as you can think. But there are so many different ways to build with Codex now. I want to show you a few of my favorite ones today. So, let's switch to a terminal for a minute. Now, we have completely refreshed the Codex TLI with tons of improvements you've been asking for. So, let's start with this empty project here in React on the right side. So, let's see if I can start voice here. We might have some voice difficulties at the moment, which is pretty unfortunate. But, it's okay. you know what we're gonna do ultra fast anyway but this time uh we'll do it with simply um the

the command line so we'll say uh build an app that picks three attendees at random i want to give them free tickets for next year so let's see what happens there so here obviously would be even faster and better with voice but we've really brought like bi-directional audio in the terminal I wish I could show it to you, but we have all of this powered by GPT Live, so the same as in the app. So this is not typically dictation. You could just react in real time with all of this. Now, a few seconds in the making here, and boom, our app is now ready with ultra fast. Wow. Now, let's do something more interesting here, and let's try to switch it up a little bit. I'll say, you know what? Make it six attendees. Why not? Don't blink. It should be fast, and boom, it's already changed. Right? So, okay. Let's bring this full screen and should we click on this button? All right, let's do it. Let's see who's coming back to Dev Day 2027 now.

All right, we have six winners. Congrats. Let's go. My dot will take it from here and email you the details later. all right so that was ultra fast and by die audio uh in the cli that all of you can try later today and we've added a ton of other new features too but let's go back to the app now i want to show you something cool here but before we do so i will actually just like start this project right now we'll go back to it in a second i want to tell you how like astra adventures came together. It's a simple idea to build a game with Astra where you could fly into space in 3D, and it all started from this sketch on paper. Now, if you take a look after a couple turns, we got someplace, but as you can tell, the graphics were a little rough. And that's where having great models across all modalities is so helpful. So as you can see, it was easy then to

kind of, after a few turns, get to a place where we have these amazing graphics and go from like a sketch to something that is very textured, very beautiful, that you can use then in a real game. So from here, all we had to do was kind of refine the game dynamics a little bit, and it should now be waiting for us. So let's take a look here. Now, instead of playing this game, what I'll say is like, show us how you play the game. What's interesting here is that now we're instantiating a browser use connection. So that means that Astra will figure out how to control the game and learn how to play it in real time. So in a second, it should be able to start the game, choose a ship on its own, and if you take a look at the screen here, on the left you'll see the decisions that the model is making, and the key it's pressing on the keyboard to kind of figure out how to play the game. It should be absolutely not an extra game. And obviously it's a very beautiful game, but I love watching Astra actually figure out a game

that was designed for us humans, but then figuring out how to play it. Pretty cool, right? So you get the idea. I'll just pause it right here. Awesome. Now, you've seen browser use, but Astra is also incredible at using any other app on your computer. In fact, let me show you. If I bring a different project over here, and let's say I bring this app on the front screen, what I can do is take an app shot. and using an app shot, I'll just dictate a quick prompt here. Audit my app across screen sizes and take screenshots along the way. I'll send this prompt. Nice. So now that I send this prompt, let's go back to the app here. I'm learning to fly. So this is just a little personal project I made for my weekend lessons to track them. Now, if you look at what's happening, I love app shots. They kind of bring all of the context, like two codecs.

And in this case, it now teaches codecs to kind of navigate the computer and in this app. So here, what you see on the screen, it's like it's starting to click on the simulator, opening the new app, figuring out the different features, also kind of opening the phone, taking screenshots and trying everything. It's really fascinating to watch live. But of course, as you're working on something else, you can just move on and let it run in the background. Okay, so now you've seen the app, the CLI, and Codex remote. I'm also really excited about how we've completely reimagined Codex in the cloud. It's fast, it works with all the same tools as Codex on your laptop, including plugins and computer use, and you can now send your most ambitious tasks to Codex in the cloud without having to keep your laptop open. You know, we've seen too many of us do that for the past year. So for this example here, let's say I'm going to create a new task, but instead of doing it with my computer, I'll just use the cloud and with this environment here, and I'll send a prompt like, rewrite the entire backend in Rust.

Let's go. Why not? It's possible. Now, this one will obviously take a while, but that's great. I know it will get done in the cloud, and I can just check on it later. I can basically send other ideas from my phone along the way and do this seamless handoff. It's been a very popular feature from so many of you, so I'm so excited that we're finally shipping it today. Everyone should just... All right, that was a lot. So zooming out of Codex, we now have incredible models across text, audio, image generation, and visual understanding. What's exciting is that we build the products like dots and Codex using the exact same tools as what we expose to you all in the API. So before I go, I wanted to show you one last demo that brings all of these modalities together. Our friends at Hugging Face kindly loaned us an early preview of their new programmable robot, MicroDuck. And that one's name is Lavender. Now, last night, I asked Codex to wire everything up here. So let's take a look if we can connect to it.

All right. Let's see now if ByDye can work over here. I love that micro doc. Look at the audience and make a cool drawing. All right, you're in this picture. Let's see what happens. It's just incredible what you can now accomplish with Codex in just a few hours. We'll give a moment, of course, for this to generate. But I added vision with Astra. I added image generation with GPT Image 2.5. I also added voice intelligence with GPT Live 1 here to interact with this little robot in real time and already give it dozens of capabilities and so much more that we could add. So let's try a couple more. Hey, Lavender, can you quack if you can hear me? Are you excited to be at Dev Day? Definitely excited. And here is the photo of all of you that he just made.

I just shared that microdust link in the chat. Check it out. That's 100 days. It's really cool. Of course, the legal doc can walk around, and Astra can teach you so many new things, but I'm running out of time to show you more. So now, the key takeaway here is that this was just a quick demo, but what I was talking about with a friend in robotics is that he was telling me he's so excited about using our APIs with hardware. With the new decision API in particular, which is able to process visual inputs, they will now be able to do things like helping robots take fast actions based on what they see in near real time. So come to me later if you want to see that in action on this little microdip. We think this will unlock a brand new way to interact with AI in the real world. What an amazing time to be a builder. You can bring your incredible ideas to life faster than ever before. So I can't wait to see what you'll do with all of these new tools. Thank you so, so much. Have a great day. And back to you, Sam. Thanks, Roman.

That was awesome. It's really amazing what you can build with these new models and these new tools. So that brings us to the third part of how we want to empower you all. We want to help people find what you build and help you build a business. Around 1.2 billion people use ChatGPT every week. It's a big platform now, and we want to enable you to build a business with it. We have tried versions of this before, admittedly with some mixed results, but I feel good about our approach this time because we are giving you the same tools we use to build ChatGPT ourselves. First, we're making it possible for people to bring their ChatGPT subscription to the products you build. So with signing with ChatGPT, they can sign into your app and use the tokens included in their ChatGPT plan. yeah that's how it should be one of the best this makes it easier for people to try your product they're already paying for an AI subscription

and now you don't have to cover their token cost to get them going we're starting with these 16 launch partners today and we're going to expand very quickly we're launching plugin extensions we are making it so that developers can now build plugins that are essentially entire applications that feel native to ChatGPT and distribute through us. With plugin extensions, you can build an editor, a dashboard, a whole workspace directly into ChatGPT and Codex. So let's look at a couple of examples. Here's our new meetings app built with plugin extensions. You can see your upcoming meetings from your calendar right inside ChatGPT. You can just click take notes in ChatGPT. The page becomes a new space where your teammates and dots can work together on the follow-up action items. Here's another example of how Figma is bringing more of its existing product into ChatGPT. You can open a design, see your team's comments, and you can ask ChatGPT to make a revision. You can work on your designs right alongside your conversation. Here's another example from Adobe, where you can use your favorite Photoshop feature right inside of ChatGPT.

Adobe and Figma are part of dozens of launch partners building new plugin extensions. We are very excited to see what you all build. And you can bring these same plugins into ChatGPT sites now, too. We've already seen millions of sites of all types get built just within a few months of launch. But now your site can support connected plugins and data. You can share a site with your team, and they can log in with their own credentials and bring their own agents. And you can see how the experience changes based on the user. It makes it easy to create personalized software experiences for anyone visiting the site. Next up, we want to make it easy for users to discover what you build. So people can search the plugin library in all the normal ways you'd expect, of course, but now we're making it so that plugins can get discovered in the flow of a conversation. When someone asks the question, thank you, ChatGPT will recognize if your plugin can help, and they can connect it right there. We've also heard your feedback about the plugin submission process,

and we're going to simplify that. you'll be able to track review, see what needs fixing, request a human review, and update your plugins tools without starting the whole submission over. So, finally, we want to make it easier to sell your products to more businesses. So today we're launching the OpenAI Marketplace. We have 30-plus launch partners today, including CodeRabbit, Notion, Vercel, and more. And our enterprise customers can now use part of their existing OpenAI commitment to buy your products. This gives you another way to reach customers already working with us And of course, if you have an open academic yourself, you can spend it in the marketplace too. Through our partnership with Base 10, you can even access open source models in the marketplace starting today. Wow. Did you hear that? Thank you. That's cool. We shipped a lot of stuff this morning. New ways to work with AI with thoughts and space. New models with GPT 6.1 Sol and UltraFast. The Decisions API, the Agents API, Codex in the Cloud,

lots of new ways to help you build a business, drive consumers to you, and a bunch of other things. It's a big day for us, and we're excited to see what you're going to do with all of this. But we wanted to end with what makes us so optimistic about this next phase of AI. And that's all of you. Every day, we hear stories about how you're using this technology to do things that last year would have felt out of reach. Working to understand the immune system. Making education more accessible. Building tools that help other people create. And we want to show you a quick video of some of these people that are actually here with us at Dev Day today. So let's meet them. Lights on. Camera's rolling. Here's to you. The builders. The new artists. The makers of our time. The momentum generation. You're not just writing code.

You're rewriting what's possible. Expanding the collection. Letting intuition guide the way. You help dream up new businesses and upcycle old ones. You're opening the doors of learning. and decoding the immune system. What you create, it lives. It changes. It moves with us. Here's to you, the builders, the makers of our time. We get to build AI every day, and that's awesome, but still, one of the coolest things about what we all get to do is to see you all succeed. It is really amazing to see what you create,

and it makes us so happy to bring you these tools. We have a great rest of the day planned. We hope you really enjoy it. Peter will be here on the main stage for a special session on how we're continuing to invest in the open-source community, including our new OpenClaw Enterprise Harness, and we have a bunch of other great sessions going deeper on what we ship today. And then at the end of the day, Tebow, Tejal, and I will be back here for a live Q&A. We're also running a Codex game studio session today where you can build your own retro game with Codex. And I'm sure a lot of you have already tried something like this, but we wanted to give you a special way to play what you build. So everyone here today is going to take home a limited edition, dead-day-only chromatic chromatic. Wow. Now for $2,000 on eBay. Yeah, pretty much. It's very cool. I think you're really going to like it. You can pick yours up in the merch shop, and you can join the Game Studio session if you want to do that. We hope you enjoy it. Okay. Finally, with all of the new things we shipped today, because we want you all to build, we figured you might need another banked reset.

There we go. So we have actually done so many of these The Tebow has been campaigning to rename us the Reset Company I think he has even made swag but more importantly I think he is out here in the audience somewhere Can we find him? Oh, there. There he is. Okay. Is he nice to clear about? That's cool. Tebow, I think it would be really nice if we could do like a low-key reset for everybody here. Like, no confetti cannons or anything. Just maybe we could do it. Can you do it? Is that a real thing? What do you think? Can we just press it? Yeah, I think so. Does it just work like that? We'll just press it and it happens. I don't know. Let's see what happens. I didn't make the scene. I am very excited. Let's do a little countdown. Ready for a countdown? Yeah. All right, let's go. Three, two, one.

Thank you for that. That's obviously for everyone in the world, not just people here at Dev Day. Before we wrap up, I want to share one more thought. People often talk about AI as a new industrial revolution. I find that quite off-putting. The best version of AI is not about making people cods in a giant machine ever worrying faster and faster. There are some parts of life that we cannot and should not automate. I believe that the future, if we get this right, can be much more like a new renaissance than a new industrial revolution. AI should be about giving people more power over their own lives, more tools to create, learn, discover, expand knowledge. And with everything we launched today and the incredible creative spirit that you all bring, we hope that we're going to get to that future. That is the future we are excited to build with you all. So thank you all very much for all you do. thank you for coming to San Francisco thank you for watching online

hope you enjoy the rest of Dev Day what do you think? first off it's so cool to watch that knowing that he is like a St. Louis kid yeah because you're from St. Louis too I always watch him through the eyes of like local boy done good yeah and it's same with jack dorsey he was st louis as well and then i always kind of had that same like and and honestly in a lot of ways similar ideals disruptive approaches um all right uh i i want all the tools i don't even know where to start grant switch off of this stream and go look at their website yeah thus far I need to

refresh my account there was a set up your dots button but it said my I wasn't ready yet supposedly oh there's an Adobe button now directly in chat GPT like just where you hit explore on the left. There's just an Adobe button, which is kind of crazy. They probably paid a lot of money for that. Yeah, I bet they did. All right, so here's Dots. So in review for folks who missed the beginning of this, Dots is their new agent tool, personal agent tool. So these are like OpenClaw, like Muse. Okay, can we talk about this for a second? Sure. Can you see this? What? Okay. I'll pause it and I'll tell you. Okay, so she's making the dots, right? So she's planning the dots. She's producing. I love that they're Mr. Potato Heads. I would buy one. This implication is that they're going to actually release this, no?

Like as a physical product, as a physical device? Or is this just like for fun? I would absolutely. I would buy that. a looks like my agent Amazon Echo kind of thing. Yeah, yeah. It just hung out here on my desk so I don't have to have my damn phone staring at it all day. You could put the face on the micro duck and have it walk around. I mean, it feels like that's like a natural place where this goes next, where you actually have a physical manifestation of this character. But immediately I saw this and I thought, oh, they're going to make a physical device version. Maybe it's just cool. I don't know. I think I would want – I'm really excited about the meta form factor, the keychain thing. And as much as that may not seem as exciting, I just think like maybe I'm in my bedroom.

Maybe I'm in my office. Maybe I'm in my living room. Maybe I'm on vacation. Do I need multiples? Do I need to drag this thing around the house with me? I kind of like the idea that I could just hook it on a belt loop and have it right there wherever I go. Wouldn't it be hilarious if the device Johnny Ive has been designing was a frog? Maybe it is. Maybe it is. We don't know. That's what I'm saying. I think it's a safe bet that it won't be. I think it would be cool if it's like a little orb thing, right? and then you could put a shell over it that's like a character. I think that would be cool. Yeah, maybe you could buy silicon skins for it that look like your guy or something. Yeah, that would be cool because then people who want something more professional will just have their little orb and people who want something a little more fun could do something like this. Yeah. I don't know. I saw this and immediately I was like, oh, they're going to actually make a physical device of it. Anything that keeps me opening my phone less is a thing that interests me.

I'll be honest. Because if I open it, it's just another chance for me to get distracted. And I will because that's who I am and that is how I roll. And that's how most of society is apparently. And, yeah, like a sheath. I really like the dots idea. You know what I want to say after watching this? When you compare this with what Mews is – what Meta is doing with Mews right now and what Anthropic has done with the models they've put out, And it feels like right this minute, Anthropic has leaned into the creativity end a little bit alongside engineering pretty specifically. Like, you know, all of the Blender stuff, all of the video stuff. OpenAI feels like they have built a tool for everyone else who works. Like, everyone is not technical. I would have said that was the opposite like six months ago, like where –

Would you? Yeah, because they launched Cowork. They were the first to do Cowork. That's true. And Cowork was a tool for everyone, right? So they sort of leaned into that direction. Sometimes the disadvantage to being first at a thing though is that whoever comes next has the ability of learning from your mistakes and the reaction to your product. and I think now I don't think OpenAI did because I think work is kind of a mess too like I love the concept but it's just all confusing it needs to be one tool you need to just go in and do the thing I don't need to have to teach people that you use one model when you hit the chat toggle and you get different models when I hit work and that on work you burn through your tokens much faster I don't want to explain that to people that's yeah it was really helping a difficult thing to teach large a large group of people like is a responsibility of grant and i and uh but but to finish my thought before i

move on i think anthropic has moved into this this very developer centric moving into the creative end of things era right now i think open ai has built something that is very knowledge worker friendly. And I think Meta may have just done something great with consumers. You still think Muse is the better consumer product? You think Muse is the better consumer product? Between that and Dots, I've got to lay hands on Dots first. Okay. I've got to lay hands on Dots first. Muse is very, very simple. But of every other agent tool you've used, do you think Muse is the best consumer product? Yeah. I really do. And I say that with the fact that its model is Muse Spark 1.3. And that's not been an issue for me, really, for most of what I'm using it for.

For what a normal person like – and I'm definitely less technical than many, many people. However, Muse is an app that if your great-grandparents decided, I wonder what these agents are and downloaded an app, they're going to be able to figure out how to set it up. If you could set up Facebook, it's harder to set up Facebook than it is to set up Muse. Yeah, and actually we've had people write in who said, yeah, I'm 71 years old, 70-plus years old, and I've been able to figure out Muse. Yeah. So that's saying something. I think it is, and I think that's where their consumer end is right now. Now, will they hang on to that? Will they edge into enterprise and everything else? But I think the tools OpenAI dropped today are smart for their position. I think it wasn't about the model at all. It's about the ability and being able to do different things with what you have,

and I think they did that very well. I think that's where we're at, and I think that's why I'm not concerned about the whole pacing the frontier idea as much as some other people are, because I don't think we need giant leaps in capabilities, intelligence, what have you, with these models. I think we need to make them incredibly easy and user-friendly to build with and integrate into everything else that we do. And, you know, my personal hobby horse, as you know, Corey, is that I really want them to be as local as possible and that we should be selling physical devices that have these things embedded into them. So then that way, you know, you're not paying the API troll toll for everything that you do. And actually, you know, we just need to rethink computing with this being the natural interface. Like it's conversational. It does things for you. You don't have to click around. You can click around if you want to, but you don't have to. Like it should just be as seamless to use as possible.

And I think we're going to get there. And I think five, ten years from now, that's not going to be a controversial idea. Everyone's just going to love that their computers are really easy to use and much easier than they are today. but it's just a matter of getting it to a point where there's no rate limits you have to worry about you don't have to worry about, someone said in the chat your voice conversations going over the cloud and someone's reading it, which I want to get to that private compute thing because that's very interesting That was going to be my follow up for you in just a second How do you feel about private inference? I think that's really an awesome approach My first thought was, oh, duh, that's a great idea. Yeah. So let's actually look into this because I didn't quite understand if it's exactly what I want. So let's read. OpenAI Private Intelligence helps businesses use Frontier AI with greater confidence that their data is protected. Zero data retention with private safety processing.

And maybe we'll switch this and open this. That's your industry standard for CDR. enables automated safety reviews without giving open AI personnel access to the underlying content. Our preview of private inference coming this fall combines confidential computing with strict verifiable controls. Okay. So then if I share this tab, it'll be a little boring and text heavy for a second. Yeah. So it enables offline automated safety review, which they just said. So it's built with three principles. So zoom this in a little so we can all read it. Customers control their content. So customer content is stored in customer controlled storage. Customers control the permissions and customer managed enterprise key management authorization required to retrieve and decrypt protected safety records. Does that make sense to you? Yeah. Yeah. I mean, that's all you want. I think that's huge for businesses.

I mean, when you start looking at enterprise concerns, like the thing that intrigues them about open source, frankly, is less this idea that they're getting a better model. It's all about this. Right now, the cost equation has not been like – at the enterprise level, the cost of serving something broadly is still quite high. Like you're still serving AI. and it's just you're putting the bill on the back end instead of the front end. So, I mean, and it'll cross a threshold where that continues to get better. It's continuing to get better now. But as the models get bigger, the cost still goes up. However, I think this is the biggest thing. Like, I mean, I've been in discussions about these things, and it's always like, you know, we're concerned. I mean, we dealt with that here at work before, like early on. And the concern was like, well, do we want to give this company access to all of our company data?

The truth is our AI is going to work a whole lot better if it has access to our company data because, like, there's a big proposition there. But if you're able to do this that way now – and the truth is you're – I don't know if it technically qualifies as ZDR, but I know that – I know it doesn't actually. But I say that. Ignore me. With your business accounts and enterprise accounts, your data is not used for training, but I would not call it private. It's still going and coming, but it is not private. This is the next level. This is very cool, and I would be willing to bet by January 1st everyone offers it. If you're going to do enterprise, you better. Yeah, this will be the standard. Yeah, 100%. So just to wrap this up, no human review, which means safety review, does not create a way for open AI personnel to read protected customer content and content retention for safety only, which means customer-controlled storage serves only approved safety purposes.

Now, I want to know what approved safety purposes are, right? Because that might be how they get around certain things. I'm going to assume, you know, no ill intent here. But I'd still want to know what that is. But that means they can't be used to train models or be made available to other groups within OpenAI or its partners. So that's interesting. I think that's about as good as it gets. like the truth is even if you're hosting it yourself at you know let's say whoever your cloud provider is I mean the truth is when it comes to major issues there are elements of access that can happen I think I think it's cool you know what I'm almost most excited to talk about is I say excited I feel like they could have said more about the $200 plan. And I get... So the $500 plan is now 25x the usage of Plus, but it also includes Ultra Fast.

So, like, if you want speed, you would get the $500 plan. There's so much stuff to dig into. It's so much stuff. Oh, here we go. New Pro Tier. Okay. Yeah, 500 with Ultra Fast. And I mean, like, you know, for a business, like a small business, I would say, and it depends on how you define small business. Small business is kind of a loaded term because, like, you can be a small business and make zero dollars, or you can be a small business and have 85 employees. So let's – Or you can have hundreds of employees and still count as a small to medium. Yeah. And, you know, it might be that for – let's say you're doing design work with AI. You might really want that speed. You're doing work that's slow, that takes a lot of time. You're doing – for me, speed is almost – I don't care. That's how I feel. I'm not in a rush where I'll pay more for it is what I'm saying.

Yeah, it's not worth it. Now, there's probably lots of instances where it is worth it to pay more, especially if you're in a highly competitive market and just being able to pay to win to get a product out faster. your competitors i could see that uh i don't like that but that's how the economy works though it's like yeah of course if you have more money you're able to do things faster uh yeah that's how it goes um but but yeah i could see using it in that context like you know hey we got to launch a personal agent before our competitors do so let's use ultra fast and code it at three times the speed um but yeah for for a normal business use case it is so not worth it to pay that extra 6x more you would not do that for like all of your employees no for me i think the 200 tier is plenty uh even with the changes like the fact is it was still at like entry pricing which was really really high like like not the cost but

what you got out of it. And I still think I'm going to go up to the $200 plan. I'm on the fence. I'm on the $100. What's holding me off right now from doing it is I want to understand DOTS pricing. Like, I want to know how much of my usage is DOTS going to eat up just by being there. because I don't want to, like, lose it. Is it showing up as available now? Can you hit the $200? Is there a – you don't even have a toggle. No, there's no toggle. Yeah, it just shows this at the moment. Okay. I go to Business and Enterprise. There's this information. I do like it. Let me see if I can go back to – Business really needed the $100 seeds. Yeah. Yeah, Peggy, that's kind of – I went from the 20 to the 100 early this year, right when it launched, and that's kind of where I stayed.

It's been really good to me. When I've hit limits, it's because I'm building things all weekend for 12 hours a day or something. Like I'm burning through some tokens, and it's kind of obnoxious. But it's had me thinking maybe I should go to the 200. And you know what? Double what I've got is a lot, especially when you factor in resets. I think it's still very much worth it. Sorry, I'm kind of half reading Grant's screen as we go here because it's all new to me too. Does anybody have other thoughts too? If anybody has thoughts, what excites you? What are you not excited about? I think the marketplace is a cool idea. I love the idea of them helping with distribution for your apps as well. Because that's the hardest part of it. I mean, it's easy now. Now you can have an idea. You can build an app, but trying to get anybody to use it, I mean, like even your friends,

It's a challenge to get people to be like, yeah, I'll go download your app and look at it. I get so many of those requests, and I'm always like, yeah, I want to. It's just a, do I have the time for dots require a business premium seat? So you've got to be on the $100 seat to use these. Oh. Yeah. I'm betting that will change with time. Probably. I'll say that. You never know, though. You never know. So, you know, the thing is, Muse is free. Like, I've been slaying Muse, and I realize I'm comparing it. This looks like a more powerful tool in very different ways. But I have yet – I've had been sitting there just waiting to put in a credit card and upgrade it to the $20 version. But thus far, I've not gone over, like, 12% usage, and I feel like I'm using it a lot. muse is free 20 and 50 i think i will say that the hundred dollar tier like if you're using it

for anything business related is really worth it it really is that i would even say the 200 tiers is worth it for business use case if it's just personal use it's not probably not probably necessary like i was talking to my father-in-law and he was like i'm not gonna pay 100 for an agent but he wants an agent that he can use. So they need to have something available at the $20 tier. But $100 for just like, I want to try it out and see if it's cool, no one's going to pay that. How does it compare to Doc compared to my agent Hermes with Hermes? Well, Peter Steinberger is the one who created Open Claw, which kind of inspired Hermes and Muse and all these other tools. So I would imagine it's pretty comparable. We haven't used it yet, though. We've used the Workspace Agents, which I assume is pretty similar to this, and it's great. I would expect it to be easier, probably. Yeah, easier to set up, you mean?

Easier to set up. Probably less like one of the biggest bugs for issues for me with both OpenClaw and with Hermes when I was trying it is, first off, you mow through tokens. You just mow through your money. I loved them. They're both great tools in their own ways. But at the same time, they're a little heavy. They're a little complex. And you've got to fix it. Like the fact is, they're kind of buggy, you know, in their own right. And I don't even mean that as a knock. The fact is the power of those kind of things unlocked is huge. and they're why tools like Grockbot and Muse and Dots and Manus' new tool exist the heck of the matter is without old Pete and his lobsters I want to comment on a couple things here

so first of all the easy to share said it's time for me to shift to Claude OpenAI is great and all, but their performance has been bad lately, and even in the demo we saw it fail. It's been frustrating not to get good quality work done by GBT. I think that's interesting, sort of like the turntables have turned. That's a community reference for anyone who gets it. What kind of work do you do? Yeah, it's a good question. You know, because I've heard this from a couple people. Much with, like, my normal work. I haven't stumbled into anything terribly bad except for the fact that I get that that's a fair comment and a fair concern. And I've had some instances when I'm using it in the web browser version. It's pretty hit or miss. Like sometimes things don't load properly.

Other times it'll time out and it'll get weird. But I'm asking it to do some pretty heavy-duty stuff. So I usually don't fault the tool for that. I've had some glitchiness in the web app of late. Not in Codex. Codex is – Codex has been enough of a gold for me. They are not prioritizing the web app. They're prioritizing the desktop app. So if you're having problems in the web app, you should switch to the desktop version and see if that improves your issues that you're having. Web development, game development. Nice. Makes sense. And are you using the desktop app or the web app? And thank you for sharing. I appreciate it. I want to talk about these two things. So I want to talk about the marketplace and sign in with ChatGPT because sign in with ChatGPT is something that I've been wanting since I've been interested in using these tools for app development, but I don't want to pay for other people's technology usage. I'm tired of buying credits. I do everything as bring your own keys with what I built, but that's really inconvenient and it's not user-friendly.

So this is a much easier pattern, and I wish everyone would update to this. It's obvious, right, that this is how this should work. And it's been coming for a little bit. But I think it goes on to the sign-in with Microsoft, sign-in with Google, sign-in with ChatGPT comes as a third level in that stack. Probably everywhere at some point. Yeah, they definitely wanted to do it for a while. And then the marketplace. What are your thoughts on this, Gordon? I think the marketplace is a great idea. I really like the ability to put things in there and have them more easily available. Sort of like GBT Store 2.0. Yeah, yeah, yeah. And I'm anxious to understand it a little better as far as what they're offering. I did hear them say open source models available in the marketplace.

And I'm curious what that looks like. how that works there's runway zendesk versell replit like i want to understand what makes this different from um like i think this is like the channel at work right like the lot of like channel we we did a whole episode on this with some of our um you know partners who cover like the the the business of the channel and how large enterprises sell things together and like package and bundle it. It seems like this is kind of like that, where you can buy a Notion subscription or an Adobe subscription directly inside ChatGPT and sort of like bundle them together. Oh, okay. I think Corey's been using Astro, so it should be great. But for me, it burns through my plus plan. Six Luna, Solver, kind of terrible release day. Yeah, I get that. And especially on the Plus plan. That makes a lot of sense.

I do – I haven't used Six Soul extensively yet. And honestly, I don't lean on Asterix super heavily. I've got an awful lot of things working through Luna. And Luna's been surprisingly good to me, partially because I'm cheap. I just – I'm trying to bleed every amount of usage I can get out of my existing budget. Power user right here. And in doing so, what I'll often do is start a chat with Astra. And then I'll be like, hey, if you need to spin up sub-agents, go use Luna or Sol for sub-agents anytime the task is clearly capable. The model is clearly capable of the task at hand. And it's done an okay job of enforcing that, best I can tell. And I do okay, but I mean I can still burn through a week's usage in a day if I'm not careful.

I just have to kind of meter it and not go too crazy. Last weekend I went too crazy. I keep the weekends. What does that mean? What did you do? It means a three billion token weekend again. And I think my highest day was like $750 million. I mean, thankfully, that's in the subscription. That's not me paying. Most of that is cash is why, you know, when you look at – when they mention things like improved prompt caching. And in Codex, there's now in the OpenAI Developers thing, this tool down here to optimize your prompt cache hit rates. anything you can do to get more in your cash as opposed to in normal input and output saves so much money so much money it's uh like they showed us the difference in ten dollars and ten cents you know there you go yeah ten cents per million tokens 95 percent less than standard input

pricing that's right yeah it really matters and that's what makes agents doable yeah because otherwise you just pay to the teeth i got the guns some people were saying like you know people who are wealthy in silicon valley and you know have cash to burn they will pay like five thousand dollars a month or something on their agents and that's just not sustainable for the average person Normal people? Yeah. Yeah. So, okay, let's say that's the pattern and they're early adopters and they want to be able to use agents all the time. And normal people, let's say the average person who is interested in this stuff, maybe would use it the same amount, right? Maybe not, maybe a little bit less, but let's say like if that's the pattern that everybody falls into, then you have to take that cost and you have to reduce that to something more like right Yeah And that like what the average person could probably do I think something they need to consider as well is I think the plan needs some attention

And I say that because most people, when you think of Earth as all people, most people using AI are still using it for free. And I think... Yeah, which is wild. If you want to increase their usage and enable them with more stuff, I think you need to make the $20 plan more attractive than it is right now. Like the truth is until five months ago, it was the best $20 in AI. I always felt like it was the deal. The Tapu Bt subscription you mean? Yeah. Yeah, the $20 chat subscription was great. Like we've got a buddy who's been building on it for ages. And on a much smaller scale than the nonsense Grant and I do, but I say nonsense. It's not nonsense. It's absolute nonsense. Until it's shipped, it's nonsense. Yeah, actually, you get the greatness. That's what it is. You've shipped a good amount of games that we need to get in the newsletter.

I have. I've shipped four games. For folks who like idle games. Yeah. We'll have to get that in the newsletter. Four games, a couple of apps, and kind of an obnoxious number of websites. Uh, many, some of which are really dumb. I'm on a kick with, uh, I'm on a, I'm on a kick for ridiculous, nothing, silly websites right now. Like, uh, I did one called the Bureau of Minor Grievances and it's just posted. It's where you can go on and just complain about stuff. Uh, and it's free, no account, no nothing. You just go slap a post it down and you're like, that watery stuff and ketchup is gross. And, uh, and it's just fun to read. And the other one is just a devastatingly annoying button. For real. Decisions API. This one's cool. Yeah. What did you think about this? I think it's Jev. You think it's Jev on the back end? No, no. I think it's just their own Jev, and I think that Jev probably should have seen that coming.

Decisions API enables real-time decision-making by focusing Luna's intelligence on a specific set of user-defined questions. with finite predefined answers, developers supply context using text or images and get back answers they can use to classify content, route requests, or choose an agent's next action. So, yeah, it's JEV. But they're using Luna to do it, which I think JEV will still win in that context. Yeah, I think so, too. I think, though, if you just want to keep them all in one bill. Yeah. Don't want to have to manage keys. like we were just talking about. Yeah, like the truth is the more you can keep in one build, the better. There was something else that I thought, oh, there goes a fleet of startups. Hang on. What was it? If you want to take over sharing, if you have something, please. Oh, no, I'm just looking at my notes. It's nothing that impressive. Plugin extensions intrigued me.

I kind of missed this. What was this? What were they really talking about here? Here's what I think. I think I think plugin extensions are just the ability to build more tools into an MCP maybe no let you give your plugin a home in chat GPT hang on I gotta widen my screen because I can't read all of your words I see yeah so what's happening is that's what Adobe is yeah you have the Adobe plugin and so it's created an Adobe page man that actually makes a ton of sense instead of opening applications on your computer you just open uh beehive and in chat gpt well because it's they're basically making it a browser at a certain level this is going to become the operating system the operating system will be this or it'll be like this will be the interface right the gui will get replaced with essentially something like this and yeah yeah that's interesting so basically yeah so you click

You go on Canva here in the sidebar, and then you have a whole Canva page. That's really cool. That's really cool. With its own conversations and projects. Look at that. Ooh. Yeah, I like that. This kind of shows you the layout. So you get your own little sidebar here. You can edit all this stuff. Oh, wow. Now I want to go and look at all of them. Get a feel for extensions with bits and bolts. Our kitchen sink plug-in. So this is the repo for people who are technical. I still think that they need, like, a really easy, like, for everything that they give out that's technical, they need a non-technical interface for it. Yeah. Well, you know, they used to. I mean, now you could just give that page to your agent and be like, I want one of these. Go do this for my business. I'm going to go eat a pizza and watch a movie. They did it. Look at them. Smart. Okay. So. Yeah, there you go. bolts. You can install the plugin.

Yeah, that's way cooler than I realized when he was explaining it. These are demo accounts to test OAuth. Your bits and bolts settings are tied to your connection in the account you choose. Okay. Also, just because it hasn't been brought up yet, a fully cloud codex is pretty baller. Oh, look at this. Help me set up bits and bolts remote with the onboarding skill. Interesting. Yeah, like that's all you would do, right? Yeah, I think so. I don't need to go read about it. Just do that and tell me what you need from me. I will provide it with you if you can't find it on my computer. Yeah, that's really interesting. So we'll let that run and we'll see what happens. I'm going to go back over here. Cool. Cool. So they're making it easier to make plug-ins, which is great. Yeah. It can host plugins. This is interesting. I don't really know why you would want this if you have your own site.

I would like to add that the automated researcher, having done some stuff and already being in use, I think, wasn't it two years ago that they said by September 26th we'll have an automated researcher? I mean, it's not made available to all of us, but I mean, it kind of is. I mean, like, I have agents. I can tell it, go do X and don't come back to you. Got it. You can use Kirk Happy's auto research repo, and you can basically have it do. I'll tell you what, his second brain is, like, one of my favorite things ever. I think people sleep on that too much. Like, it is so easy for me to now sit down at a new agent and be like, go read my Obsidian and come back when you know things. I missed this. There's shareable profiles. profiles that people showcase sites and plugins that others can find and reuse a profile brings those creations together so you can share them in one place teammates can also discover shared skills with skills sharing limited to the workspace

available to all business and enterprise so basically you can have a cory page and you can have all of the plugins that you've created and all of the sites that you manage and it can all be in one place so that people are like what's cory working on and they can go click on you and see all your cool stuff. Kind of like my GitHub front page is now after our recent tutorial with... Oh, my God. Oh, with Cassidy from GitHub. Cassidy. I literally went home and did everything she said. And I now have a nifty homepage with links to all of my stuff. And am using GitHub much more effectively than I was before. Which was absolutely the goal in reaching out and asking for someone Check this out. Look at this. So this is my profile. It's already built in here. Lifetime tokens. This feels low. 248 million. That feels low.

60 million peak tokens. Longest task, 29 minutes. I think some of this didn't count further back. Yeah, there must be some. That doesn't make sense on mine. But can I show mine? Yeah, go ahead. I'll stop sharing. Oh, hang on. I hit the wrong thing because I'm going to do this. There we go. As you can see, that's my test account, so it's not my main account. I was pretty slow last week. I've been mowing through my rate limit too fast on the weekends. You know what? It also doesn't cover workspace agents where a lot of my token usage was happening, so that's probably why. Yeah. Okay, let's see yours. Yeah, $12.3 billion. Putting me to game, bro. What the heck? Yeah, like, look at... Wow. $394, $300 million, $258, $418, $240, $360, $743.

That's my day to beat. $743 million tokens. But I like shipped a mobile game and updated three over that weekend. you know uh yeah this is on your max plan is that what this is yeah no this is my yeah my hundred dollar plan yeah right pro sorry pro for this and and i've decided it's not enough so i need to go to but like you go way back here and i mean that's higher than i expected and i think it's only counting codex yeah so like these days that are empty or low that doesn't mean i haven't been doing like a mountain of stuff and uh and other things oh it locks everything down here too like see fast mode zero yeah mine's been 65 for fast mode i was like what are you smoking these are my biggies right here product design oh yeah computer use chrome yeah nice nice yeah like i use those heavily total chats 4,875 i don't know if that's a lot but it feels like a lot

wow this is quite impressive i wonder what the highest one is they should make like a public leaderboard so that everyone can compare it's vincent kotch from open claw that's who it is it's 54 billion token day but come on there's got to be people inside open ai who do more than that the comments on on that on the short of that of him saying that are crazy yeah beasy 12 billion isn't that wild I'm telling you I got stuff going every day well less so the last two weeks I've had some stuff going on I'm dealing with and this stretch right here that you see mowed through every rate limit I had every reset weekly so you can look at them weekly and yeah you can see 2.8 billion token week wow you should honestly like you need to run the numbers on how much money this

is equivalent to an api credits now that they've changed it like you should just look back like what's the percentage of end out yeah like and i would love to know the caching percentage yeah artificial analysis does like a blended rate um yeah so it's sort of i'd be happy to know what their ratio is. Ask ChatTBT right now. Be like, hey, just copy and paste all of those numbers and say how much do you think this was? 19.8 billion. That's higher than you. Hey, Johnny, what is what ratio does artificial analysis use when they are trying to decide how many input, output, and cash tokens to use as a comparison. I worded that poorly, but I hope you know what I mean. Okay, unlikely lab in the chat has the highest so far

at 19.8 billion lifetime tokens. Does anyone have higher than 19.8 billion? We can't hear your audio. No, I know. It's because I shared the full screen. It's my other user profile, so I can share my personal one. And he says 70% cash hits, 20% input, 10% output. Okay. On 12.3, 12.8 billion. I can't remember. Oh, and I'm already moved out of it. we're going to go with 12 point. So BZ has 1 billion tokens. Nice. Oh, Gene Patrick Smith in the chat. I have like 200 billion, LOL, probably more. Wow. One of my accounts has 39 billion. Okay, John Patrick. I'm assuming it's John Patrick versus Gene Patrick.

You need to send us a screenshot. You need to host it somewhere so we can see the screenshot. No joke, man. This is wild. I love it. $200 billion. Is that on the – what plan is that on? Grant, what model do you think I should use for it to base that on? I'm going to have it base it on 6.1. It's smart. Tell it to do a spreadsheet and track, like, when all the models came out. You were probably using the best ones at the time. Hang on. Just have it do that for you. Add credits. That's not a good sign. Hold on. Are you out of usage at the moment? You can do this on your work once. You don't have to waste your personal time. Oh! No! 292 credits remaining. Oh, you could do... Yeah, those don't go far, though. That's misleading. Like, one Astro runs, like, 250 credits or something.

No joke. When I first started buying these back in February, $40 got you 600 queries, and it was queries. Can you put this comment up on the chat? This is important. Yasser says, can we talk about how dots and most stuff is not allowed in Europe and the UK? That's a really good call out. That is a really good call out. What is happening? Yeah, Astray in the chat said you do have a reset, Corey, you can use. Yeah, one came. Thank you. Good catch. Yeah. I appreciate that. For all of our friends in the UK and Europe, I'm sorry. The regulations are restricting you here, I believe. Or they just didn't rule it out to you because of the restrictions. Yeah, that's been a really ongoing issue, And it really sucks for Europeans and some other countries that it's really unfortunate because I know that there I have European friends who are always like living in a state of FOMO from all the cool stuff that for one reason.

I interest you in a VPN and a virtual machine. Yeah. Like, I think you got to just work around that. Yeah. It's about all time to move. I wonder how well VPNs do stay around every day. But I imagine it does because I think there's a lot of people who probably are working around these limits. Yeah. John Patrick Smith says, I have like 10 pro accounts. Nice. I don't know how you pulled that off. I don't know how many phone numbers you have. That explains it. Yeah. Yeah, but opening, okay, Sethi Pagna says, OpenAI plans usage limit is flashed 50% of what it used to be official. Yeah, yeah, that's for the Pro account, I think. Yeah. That's right. And then. Yeah, there was some word salad around it to do with how it'll actually be the same amount or a little more usage when you saw all the other stuff coming.

I don't know if there is a cut in the cashing price. There may be some other things. I don't fully understand, but I'm hoping to keep an eye on Twitter over the next day or so and see someone come on and kind of explain the math a little more clearly behind that. I'm probably still going to upgrade to the $200 plan. You're going to do it. Yeah. Well, I mean, it's still double what I've got. I mean, it's double the money, but it's also double the usage. Right. Which seems fair. I mean, like, don't get me wrong. I would much rather have it at the 20. Also, if you currently have the 20X plans, you will keep those for a while. They said they'll be phased out at some point. But for the foreseeable future, it looks like if you have it now, you're okay. It's just for new upgrades right at the moment. so um a couple comments in the chat here um someone said uh i have two android phones

or this is g Patrick again uh i'm canceling all this uh opus and sonnet are good enough for me keep one codex that's where i'm at where i have one codex and one um yeah max cloud code plan and that's pretty good um and then you know if you need to pay more than that probably better going to a third-party, like, open code or something. Yeah, don't sleep on open code. Yeah, it's pretty good. Like, I mean, there's even free models. Like, you know, I'm using NewSpark 1.3 and one of the DeepSeeks through there for free. They got the Nemotron models. So, like, it depends on what you're doing. It's not necessarily the cutting edge of greatest intelligence this week, but it's still better intelligence than was cutting edge 90 days ago. Right. For sure, yeah. 90 days ago, this level was just fine. So I think it's important to remember that depending on what you're doing, you really may not have to spend the big money.

Like, I'm getting great use out of lesser models than what I used to. Omo. I don't know. yeah almost we've seen before in passing open code 2 i didn't realize i had a 2 or maybe we shared that a week or two ago and i just haven't tested it since then it may be i'm still running open code the desktop version and and it's it's been good to me i spend a little bit of time with um a little bit of time just in open router noodling around through different models uh I didn't used to be good about that, and I've tried to – like making sure I'm taking a little time every week to just go try kind of what's new. It's where I found that space bunny model. I need to do that more too, yeah. Like we'll cover stuff that comes out and test it in that interim, but then I just kind of abandon it and use whatever I think is the best or the most convenient for me. Yeah, convenience is worth a lot.

I think we will continue to see convenience become a bigger factor in the AI people choose to use. Because when you get into this next wave of users that's going to continue coming, like, not everybody is going to walk in the door being like, oh, my God, we got two more points on DeepSuite. They're not going to care. Who cares about that? That's the thing. I don't care about that. Yeah. Yeah, it only helps you if you're saying, like, can it do, like, broad strokes what I'm trying to get it to do? And is it the best one at that? But the actual numbers and all of the shenanigans about it, it's like it really doesn't. It's not helpful. Like, you've got to just try it for what you're trying to do. It's good. It's marketing, like you said. Like, it's really good at putting it in front of you and say, hey, this model actually might be really good for what you're looking for. And that's going to become incredibly important. Brudge, grudge. As we go into like a multi-model world where a lot of people are going to be training their own models and those are going to be disintermediated by agents.

So what's going to happen is you're going to have custom agents that you're interacting with more so than models. That's my personal belief system where at a certain point you're not going to even know what model is running under the hood most of the time because it's going to be agent to agent communication. so in that world it is important to know what specialist agents like chat dpt was talking about is the best at what you're trying to do or best meaning most intelligent and cost effective my one caveat to that I 100% agree I guess I 99% agree I have a caveat that makes it not 100 yeah you don't 100% agree well it's not disagreeing. It's more of a tacked on point is that I think it's really, really important that we kind of ignore the benchmarks shared by the model companies themselves. Like if Anthropic or OpenAI or XAI or anybody, if they're sharing it, you should expect that it is

heavily marketing speaked up. Like you'll notice and both of them do this incessantly. Each each benchmark shared on the 6.1 soul page as you go down you'll notice some of them show fables some of them show opus 5.5 well why do the others not yeah sometimes it's because that one didn't take it other times it's because they're like well this doesn't you know further our marketing story and same goes for clod i've seen that i've seen anthropic do it a hundred times too and there's always very selective placement and choosing of the benchmarks. So would someone like artificial analysis? I like artificial analysis. They're probably my go-to. And I also like going to like LM Arena. While I don't think it's perfectly factual, it is – Captures maybe more of the models. Your reaction. Yeah. There's a smell check by real people to that.

and I think combining that with what you see on like an artificial analysis or one of the other independent numbers like I think you get a better idea of what you're looking for before you go strapping something to your expensive workflow I think if we're going to do by the way two seconds if you haven't yet please take just a minute to like and subscribe while we have a whole bunch of you on here we heard there's some new people who found our channel for the first time and you like it so please subscribe so you'll see yeah when we do stuff we're we're glad you're here and we're Glad to spend a lot of fun together. Okay, so if we're going to go down this regulation or industry standardization path, which makes sense to where we're at in the industry at this point, then we should standardize the benchmarking and the evaluations. Now, that doesn't mean that we should not add new evaluations to the standardization, but there should be some sort of, like, formal, like, reports that you give out. because they all do system cards, but those system cards should be standardized so that everybody follows the same rules

and you don't get to leave out things about X, Y, Z topics or risks or hallucinations or whatever. We didn't do good on these. Well, just don't put them in the press release then. Yeah, yeah, yeah. No, you need to put everything in there, and everyone needs to run the same tests, okay? Yeah. And you need to have the same rules for the tests and you need to have the same features and the same harness for the tests so that everybody is on the same page and it's like very standardized. I think that's a good place for us to go from here. Yeah, I think when you look at these and you see that they're like, oh, look, they shared five benchmarks. I want to see the 30 you didn't share. Show me what you ran these through that you didn't decide was press release worthy. I would love to see those numbers. And I think you're right. I think just a reporting structure, not even so much a full regulatory thing, But like, hey, yeah, don't need that. You know, we're going to run these 18 benchmarks and no matter what the results are, we're going to publish them. And you can use those alongside others. And and it would give a lot more respect to them.

I mean, I think you could still highlight, like, look how good we did here. I think it's fine. Yeah. Yeah. The truth is, I'm always going to trust it more from a third party anyways, probably. Sure. That's fair. I just just because marketing. yeah yeah but i do think you know everyone's everyone's going to want to customize their own models under the hood because they want to control costs and i think everybody's going to try to create an agent he wrote a whole thing about this uh in today's newsletter about like the personal agent aggregation land grab right which is everybody wants to be the personal agent that interfaces with everything else um and the the delicious word poly agent agent morris so if you're familiar with the term polyamorous, polyagentorous, I think is a cleaner way to say it, where you will have a multiple agent world and your agent will talk to other agents and you might have multiple agents that you're in relationship with and that you're talking to. You're chief of staff. Yeah, but there is still going to be an aggregation effect. There already is, right? We're

already seeing an aggregation effect where most people go to ChatGPT or Claude or Gemini or Muse now to get information as their window in. The browser is getting consumed by the agent window. So in that world, right, you are either an agent competing with all of the big tech, like hyperscalers, like most valuable companies in the world, or you are a service provider that is now getting, you know, you're going to be aggregated in the feed and your service better be the best or the agent's going to pick something else, right? So it's a very scary world that we're coming into. It is. I want to get to a couple of these last two questions before we have to cut because we've got to go pretty quick. Let's do that. Do you think Dots will be able to act as an ongoing SEO agent? I've got just a tiny itch as to who this is. I think so. I get the vibe you'll be able to build up different agents as you want them to be, give them different instructions and tool access. So I think that will be the case whenever it's ready. I'm anxious to see how quick

it gets to my particular flavor of chat GPT plan. But we'll be keeping an eye on it, and you can bet as soon as it's around, Grant and I will be showing up, hopefully by Thursday. I would love for all of this to be, at least some of this, to be available on Thursday to be able to walk through and demo. It seems a lot of it is business plan restricted at the moment. Yeah, and a good number of business things, which is good too, because they don't always get as much attention either. When would 6.1.Soul be available? Artificial analysis benchmark gives me hope. Yeah, same. thing uh sole 6.1 i don't have yet that doesn't mean that it doesn't exist yet uh but i don't recall it saying anything about a delayed release but it's pretty common for when models come out for us to be a few hours so we're on artificial analysis.ai oh there we go see this is 6.1 max here oh that's great meaning max effort yeah and then six gpc6 astra max so they're just about

neck and neck that's great for a fifth of the money yeah you know close enough that it's uh that it's how big a jump is that that's a five point jump over five six soul which i'm still using for most of my chat to bt things most of my stuff is still running on it I finally switched them this weekend. Okay, tell me about that. Yeah, I had them all switched this weekend because they cut the price of 6-Sol down to significantly better than 5.6-Sol. So I was like, well, I was just thinking in terms of eating my bandwidth, trying to keep as much as I could. All right. What is that 5 preview? quick note step 5 preview interesting what the hell is that it's in the mix here yeah we'll have to do some googling yeah oh I think I think I might have covered this Space Bunny

no no Space Bunny is probably going to bind up being Mini Max it looks like yeah I don't see it there okay this is their blended by the way just to close the loop on this we talked about input output blended rate this is what they call it so So $0.72 toss per intelligence index task. Cool. Yeah, and it said their figure was 70% cached, 20% input, 10% output, which sounds pretty accurate to me. Yeah. And just for comparison, so if we compare it to 5.5, it's expensive. It is. Yeah. But, you know, it's smart. Yeah. If that extra bit of intelligence or extra bit, depending on the tools you use, matters, then it's what matters. I'm grateful that for me it doesn't. Yeah, for what you're working on, it's not an issue.

It's just not an issue for me yet. I mean, don't get me wrong. I'm not against better tools, but I'm pretty cozy where I'm at with what I get out of what I pay. But I'm still going to double it. Well, guys, to everyone watching, thank you so much again for joining. Please pop by the Neuron.ai, sign up for the newsletter. We'd love to count you as one of our readers. Thank you for joining. Thank you for chatting. This is so much fun, and we absolutely love doing it. We'll be back on Thursday at 12 p.m. Central, 10 a.m. Pacific. So if you're out and about, come join us. Probably we're going to wind up poking through some of these tools, whatever we have access to yet. So otherwise, you know, bring some questions, bring some thoughts, whatever. we're down we're down to clown see you on Thursday farewell for now farewell for now oh sorry yeah I blew it I blew the whole