¶1Here's everything that happened at OpenAI Dev Day, which I was fortunate enough to attend. There were a lot of announcements, many of which are very cool, including everybody getting one of these, a mod retro that you can literally customize and build your own games, of course, with codecs. All right. So, probably the biggest announcement of the day was DOTS, OpenAI's new feature to compete with Grockbot, Muse, Instinct, all of these new generation of highly personal AI assistants, all inspired by OpenClaw. So, shout out to Peter Steinberger and OpenClaw for basically setting the tone for the entire next generation of personal assistants.
¶2You can think of Dots as always on. It is always working for you 24 hours a day. It is much more proactive. It is not something where you're going into chat GPT typing a prompt and then getting a task done. It is much more of it trying to think of what you need done before you actually need it.
¶3And you can connect to it through chat GPT obviously, but also Slack and Teams. It is where you need it. Again, I've been using Grockbot a ton. That is my primary personal AI assistant right now, but of course I'm going to give DOT a try. Now, I think the one interesting thing is it seems like all of these personal AI assistants are converging around the same design language.
¶4Some kind of soft geometric shape with eyeballs and that's it. And you can see both Grockbot and Dots have very similar design language. It feels very familiar and it feels just like what happened when Chad GPT was released. Every single product started looking like the Chad GPT interface. So just like Grockbot, it is its own agent or set of agents working as dots.
¶5They get their own environment working completely in the cloud. You connect it to your personal accounts like Gmail or Google Docs or Calendar or whatever you're using. You connect it and then it has access to all of it. I really am excited about this because it seems like this is a good way to get the general perception of AI to skew more positive than it has been. When people get this simplified product which is not like codeex or chaibt it is really it just feels like you're talking to an AI assistant and that AI assistant is actually getting real work done for you.
¶6I feel like it's going to convert a lot of people into actually enjoying using AI or at least being okay with it because it's just getting so much done for you. Now the way that DOT works is kind of interesting. It lives within the chat GPT app and it actually controls your chat GPT and codeex threads which is really a very different approach than what Muse took certainly but also how Grockbot works. And to be honest, I don't love this approach. I kind of wish DOT was just its own app, its own thing, but I guess for it to have computer control and be able to code and have its own environment, they just built it right on top of codecs and chat gpt.
¶7I get it. Now conversations with dot don't count towards your chat GPT usage which is quite nice but as soon as you have dot control your chat GPT threads or your codeex threads then that counts towards your usage. So just keep that in mind. Okay. The next one and one that I am personally most excited about is ultraast.
¶8This is GPT6 Astra powered by Cerebrus chips and it goes eight times faster than what Astra would do normally on their Nvidia GPUs, but it comes at a severe cost. You are paying six times more. And let me tell you, those tokens burn quickly. I had early access and I burned through $1,000 in literally like an hour and a half. So, you can really burn tokens very, very quickly.
¶9So, be careful with that. And with that, they did introduce a new plan, Pro500. That is a $500 a month plan. Now, I know a lot of people hear that and they're going to think, "Oh, everybody's saying AI is getting cheaper and cheaper, but people are paying more for it." Well, the same intelligence that is available today for pennies per million output tokens were, you know, 10, 20, $30 just a year ago for the same exact million tokens. So, intelligence is decreasing in price, but the absolute frontier, you're definitely paying more, whether that's for speed or just how high quality those tokens are.
¶10All right, so here's an example of how fast ultraast is, and it it's crazy fast. So, here's standard on the left, ultraast on the right, and they're both going to be building rockets and launching them. Okay, so you can see standard going pretty good. Ultraast building already getting started on the rocket. And while standard is still thinking, ultra fast rocket is already launching.
¶11So with that pro-500 plan that I mentioned, you have higher usage limits. You get 25 times more than plus. So the $200 a month plan, which was 20x, is now only 10x. The $500 a month plan is now 25x. And what that means is you're barely getting more than what you used to get at the $200 mark, but now you're going to be paying a lot more.
¶12Okay. And possibly the most impressive announcement of today is GPT6.1 Soul. This is another incredible model that is affordable, fast, efficient, and basically as good as the Absolute Frontier. We saw Enthropic release Opus 5.5, which was the best model in the world, or actually still currently is the best model in the world, but significantly less expensive than Fable. Then they announced Sonnet 5.5, again, basically as good as Opus, half the price, and now we're getting the same kind of releases from OpenAI.
¶13GPT 6.1 Soul is a phenomenal model. It is basically as good as Astra and much cheaper and faster. So GPT6 Astra is coming in at $10 per million input, $50 per million output, 6.1 soul, $2 per million input, $10 per million output, 10 cents for a cash million input tokens. That is so cheap. It is so so cheap.
¶14And the model's really good. Let me show you the benchmarks quickly. First of all, the probably best benchmark that I know of right now, best reflecting how actual developers feel about these models. Here's 6.1 Soul right there at the same level, if not higher than Astra, and it's a fraction of the price. Look at how good that is.
¶15It seems like Anthropic and OpenAI both figured something out at the same time where their post-training is basically just solved, right? They build these massive models, Fable and Astra, and then they are able to post-train them, whether they're using distillation or whatever techniques they're using, into a model that is smaller and basically as good, much easier to run inference on, which means cheaper prices, faster inference, and it's good for all of us because we get these phenomenal models. Here it is on GDP PDF, which I assume is a PDF benchmark or PDF creation. And you can see the same thing. It basically performs like Astra at a fraction of the price.
¶16Here it is with OS World, which is computer use. Once again, performing very well, especially at the highest thinking tiers, but Astra does at the very top end perform better overall. But look at that. You're paying just a fraction of the price. All right.
¶17Quickly, they also released Codeex Security Cloud, which is exactly what it sounds like. It is AI that constantly scans your code for any issues, vulnerabilities, and just continuously tells you if there are any issues with your code. Next, Codeex is going to the cloud natively. Now, Codeex belongs in the cloud. This is the way things should be.
¶18You should be coding in the cloud. You should not have to have your laptop open at all times. That never made sense. I am very excited about this. Plus, it seems like the biggest bottleneck now is my local computer, especially when I have ultra fast mode going.
¶19Let me tell you, the actual inference is lightning fast. And then when it's doing tool calling or if it's running terminal commands, that's the slow part now. It's kind of insane to see. Ultraast really exposes the bottleneck is now your computer. They also have a refresh CLI which allows you to control it via voice.
¶20They have a new code review experience in codecs and watch out Jev. They now have a preview of the decisions API which is a model built to make quick decisions just like Jev. Now it is interesting. They say it's based on Luna's intelligence. So it is their lightest, smallest model running super fast making decisions.
¶21But I still think Jev is going to be faster. that is a model built from the ground up to be a decision model rather than repurposing an existing model of Luna to do this new thing. They also relaunched plugins where if you remember about two dev days ago, everybody said apps are dead or here's how many startups that OpenAI just killed. That entire experience did not work very well and they are relaunching it. Now you can have apps natively in the chatbt experience and you can also log in with an open AI account sign in with chatbt to third parties as well and you can bring your tokens with you which I think is very generous of them.
¶22Doesn't seem like something anthropic would do but this is really cool. So imagine you already have a Chad GPT subscription. you go and log in to some third party, some other app that you use and rather than that third party having to charge you an additional subscription or provide you with tokens, you simply bring your tokens to them which I think is going to be awesome. And then watch out notion chat GPT space is here. Think of it like notion but it is fully native for you and your teammates to work with and with agents to be in there.
¶23Everything from PowerPoint to spreadsheets to websites. Everything can be in this new chat space. So, even with all of those exciting releases, this is what I'm most excited