A year ago, ChatGPT felt untouchable. Not anymore.

Mm-hmm!

Hi everyone. It is Thursday, February 5th, and I'm Deirdre Bausa with another Tech Check livestream. You might remember if you tuned in a few weeks ago, we sat down with the Replit CEO Amjad Massad and talked about this AI coding revolution, how people like me and producer Jasmine with zero technical background are suddenly building and shipping real apps. I'm not saying they're any good, at least mine anyways, but we are building them. That itself felt like a big moment. But this week's, even this day, makes that feel like it was just a warmup. Uh, since then, the software sell-off, it has only accelerated and it has spread. It started with project management tools and CRM. Now it's hitting legal tech, video game developers, financial data companies. JP Morgan out with a note this week, and they put it this way, it's a broad sweeping mass exodus of software stocks. And let me put it another way, it has been mostly a bloodbath. Now, also over the last few weeks, we've had OpenClaw. You might have heard of it. It's an open-source AI agent built by just one guy, and it has gone absolutely viral. Essentially, you hook it up to your computer and it— give it full access and just let it do things, act, build, book things on your behalf. Uh, OpenAI meanwhile has launched Codex, a new version of it, its answer to Claude Code. Altman says that Codex is the company's most loved internal product ever, but somehow, at least it feels like to us, it seems like it's still struggling to get the same kind of buzz that Claude Code has been able to achieve, especially over the last few weeks. And by the way, nobody seems to be talking about Gemini 3 anymore, but as we know, that could change very quickly. What else has happened? Anthropic has really pounced on this moment yesterday, dropping that Super Bowl ad, which I hope everyone has seen. Taking direct shots at OpenAI for putting ads in ChatGPT. You had Altman firing back on X, calling Anthropic authoritarian, and this whole debate over elitism in AI. And then today, guys, both companies, they have dropped major announcements. Anthropic released Opus 4.6, their most powerful model yet. It can work directly inside PowerPoint. It can pull together regulatory filings, run teams of agents in parallel, OpenAI countering with Frontier, a new enterprise platform to build and manage AI agents. The pitch for them is that they can do real work across your entire business. Okay, so that's now. But a year ago, ChatGPT felt untouchable. Now it is a full-on war for the enterprise, for the developer, for the consumer. So we wanted to go deeper and Guillermo Rauch, he is co-founder and CEO of Vercel, one of the most important infrastructure companies in tech. Right now, last valued at more than $9 billion. If you have used ChatGPT, you have used Vercel. He's been building the AI tools that are changing how developers work, and he works with both OpenAI and Anthropic. He is really the perfect person to help us make sense of it all. Guillermo, thank you so much for joining us.

Thanks for having me. It's exciting to be here, especially this week. There's so much happening with AI.

The last week, the last few weeks, I mean, Jasmine and I were bystanders, we're onlookers. I can't imagine what what it's been like to be on the inside. We had Lisa Su, the CEO of AMD, on our air yesterday, and she said that she's also been blown away by the last 60 to 90 days. You couldn't be more of an insider as well. Yeah. So what are you seeing? Is this acceleration, is the hype all real?

Yeah. To just give you a picture, right? Vercel is a place where developers host a lot of the new applications they build and deploy to the cloud. So whenever you have an idea, Vercel makes it so easy to just make it real. And increasingly, you know, when we started 10 years ago, you needed to have hard programming skills. What happened during the Christmas season, like just like the, the week of Christmas, the week of New Year's, every person that I know was playing with Opus 4.5, vibe coding and bringing new applications live. We saw a massive spike in new deployments on the platform, which happened at a time where you don't expect a ton of activity. Like people are You know, going home for the holidays. But what ended up happening is everyone went home to vibe code and we saw a 14 times larger growth from November to December than we saw last year. So what's happening is software is now effectively free and, you know, we can sort of like link that back to what we're seeing in the stock market, but all software that we're used to been buying, installing, configuring is now being automated. Being generated by these models. Yeah. And Brazil is in this awesome position where you have to deploy it somewhere. So, you know, there's work.

And you're seeing it and benefiting from it. And Guillermo, I'm with you on this, right? I'm one of those people. I would say it was a few weeks after Christmas that Jasmine and I started really getting into it first with Cloud Code, then Replit, Lovable. We're, you know, plans to test out Codex as well. It's funny you say that. Maybe next year everyone's gonna be asking for Cloud credits for Christmas. But this moment now, I mean, I've been, posting about this, experimenting about it myself. Jasmine and I, just a few days ago, we created our version of monday.com. We woke up and we're like, wouldn't it be fun if we could just try and do this? And an hour later, we didn't have something great, but we had something usable. And we just thought like, wow, if we could do this and we have no coding experience, what could someone who does or someone who knows enterprise software actually do? That said, I have been getting a lot of pushback, right? People say, no, you're not about to go in and do this. It's not secure. You don't have the database. SaaS is not dead.

Yeah, and I'll tell you why, right? So every week, these models are getting better and better and better. It seemed like vibe coding was going to be a temporary thing because a lot of what you could build looked like prototypes. But I'll say, ever since the release of Opus 4.5, there's emerging consensus that you can go further. Each time a new agent or a new model drops, we can go from prototype to production-grade applications. And the, the image that I want to paint for people is when you deploy something that you've vibe-coded with an agent like Cloud Code or v0 on Vercel, you're actually using the same cloud infrastructure that serves some of the largest properties on the internet. It's using the same frameworks like Next.js, it's using the same global cloud infrastructure, same security layer. So vibe coding is sort of going through its teenage years into adulthood very, very quickly. And now truly everybody can create software.

So I guess I asked this question directly of Amjad a few weeks ago. I'm going to do the same for you. Is software facing this existential crisis? That's interesting what you brought up. You're saying that it actually is maybe more secure than people think. Um, do they all— not all of them, or at least some of them, or how many of them go to zero? These public companies that have, you know, $5 billion plus market caps, they that people can now vibe code.

Yeah, I won't be as extreme as saying that all software stocks are gonna tank. But what I'll say is there's gonna be a lot more just-in-time software. There's gonna be a lot more personal software. There's gonna be a lot more throwaway software. And this is basically enabled by the fluidity of platforms like Vercel because anyone with an idea within the enterprise, the data already exists. It's in the sources of truth, right? It might be in Salesforce. It might be in your Snowflake data warehouse. And now the UI layer with Vercel technology can be fully generative. You can bring an application to life for the problem that you have that very moment. So you don't have to buy this like sort of static and terse software that you have to configure to your needs and tends to be really expensive. So I do expect a downward pricing pressure of— and, you know, this idea of like month-long procurement cycles for what is now effectively legacy software when you can just prompt it into ChatGPT and deploy it on Vercel with basically two prompts.

And certainly if you don't think it's gonna go away, that's gonna have an impact on pricing power. The way I like to describe it, because I'm one of those people that have just been experimenting with it, Jasmine as well, and we can't believe that everyone's not doing this, but it, like, the AI race has gone from talking to your chatbot to your agent doing things for you. And like, Maybe that doesn't get across because we've been talking about agents for a year already. But what I'm seeing, and I think what you're saying, is that like ordinary people are now able to use it. So with that, talk to me about Opus 4.6, if it's a game changer. I know you've been texting with a bunch of founders and CEOs in the AI space and OpenClaw too, one of the most viral moments of this new era.

Yeah, the really special thing about Anthropic's models is just how good they are at this. Agentic loops. When models can use tools, they gain superpowers. The reason that agents are getting really good at programming is not because they're like these geniuses that immediately spit out the perfect code. It's because they actually can hold tools just like us human engineers do. For example, take the Next.js framework, which Vercel has created. It can go back and forth and read the errors and iterate through it and basically work like a human being would. And so you kind of can extrapolate this from, you know, every expert human job, every knowledge worker. There's a series of steps and tools that we use that can be replicated by these agents. And I think Anthropic has really run away so far with this idea of how to automate. Instead of coding, you can think of it as automating the operating system. Imagine that under the hood, the agent is using a computer just like your engineer would, your designer would, your marketer would, and then it gives you the output, which is what you want, right? And I think it's slightly different from how we typically think about coding models. Codex by OpenAI is an extremely powerful coding model. In fact, you were mentioning ClaudeBot. This project sort of seemingly has come out of nowhere and now is giving everyone their own Jarvis. Assistant is how I explain it. It's automating computing, and it's created by one guy.

Iron Man reference, right? That's right, Tony Stark, Iron Man reference. Okay, making sure.

Yeah, yeah, go take care of my laundry. Hey, go open the door for me. Hey, do this, cook me a meal. So we're kind of approaching this world, or the holodeck of Star Trek, right, where there's just ambient computing everywhere and you can use the conversational interface to literally do anything. Okay.

I want to get more into, I want to get more into Open Cloud, but first I just noticed that you talked about Anthropic being really special and then you talked a little bit about OpenAI's Codex, but I mentioned this sort of at the top. It feels like Anthropic has stolen all the buzz, not just among developers, but it feels like among media, maybe mainstream. Can you compare for me Claude Code versus OpenAI's Codex and Is OpenAI on the back foot right now?

Yeah, I'll say, look, Anthropic is really good at this task of automating your computer tool calling. Codex and OpenAI seem to be really good at expert software engineering tasks. So clearly they both have their strengths and weaknesses. Within Vercel, I'm seeing that our most capable agentic engineers use a combination of models and tools. You might have the the first idea, you go to v0. You might go to ChatGPT. You might, when you're doing research, you go to ChatGPT Pro. So the reality is that we are in a world where the winners really are the developers and end users. I'll say this too. Anthropic was winning in the hearts and minds of developers. I think when it comes to that ad, you know, was it necessary to bring the comparison back to a consumer product in ads? I'm not sure. And at the same time, OpenAI is realizing that nailing coding might be the most— yeah, I believe coding is the killer app of AI. I believe that's how we get to artificial superintelligence. So it's a good time for both companies.

By the way, we're reading the comments, so keep sending them in on YouTube. But okay, you just said something like, why did Anthropic have to go after after OpenAI sort of on their strength, right, on the consumer side. But if you believe that coding is like the killer use case, does that make sense for Anthropic to do? Is Anthropic fighting OpenAI not just in the enterprise, but could you see that happening in consumer as well?

Well, I think ClaudeBot is a good example. Like, imagine that every assistant and every, uh, magic AI tool that you're talking to under the hood it actually has to use something like Code and Model, which is actually kind of interesting to think about, right? Like it's almost like—

Guillermo, can you explain OpenClaw too?

Can you explain it to people that might not be familiar? Yeah, it really is a huge disruption in how we interface with computers because every time I need to automate something, write a document, et cetera, I go to my computer and I launch applications and I, if I'm more technical, I run commands and I write some code. Now take a step back and imagine that you're not gonna use the computer directly. Let's let the agent use the computer and I interface with the agent over WhatsApp or Telegram. So this is how ClaudeBot, or it got later renamed to MaltBot, works. You talk to it like you would talk to a friend or your assistant over text message and you tell it, what are emails I should pay attention to? I'll give you an example that I used 'cause I built my own little ClaudeBot agent like this on Vercel. I wanted to unsubscribe from every spam newsletter. We have all this amazing AI and I keep getting spam and things that I didn't sign up for. So I told my assistant, just go through all my email and hit unsubscribe for all the spam. This is what technologies like ClaudeBot can use. For the user, it's the most user-friendly computing device in history because you just talk to it. Under the hood, believe it or not, it's using these really advanced coding models. It's almost like having a super intelligent engineer at your disposal that you're talking to over text message and understands everything you wanna do.

Right, I've heard it been called AI with arms too, just in the sense you're right, you don't need to be at your computer, you can give it directions. That is also, to me at least, a very scary thought. I'm already uncomfortable with giving—

Go ahead, sorry. So, Rafael launched a product called Sandbox, and as the name implies, we can put these agents in little sandboxes where we give them freedom. Because I'll tell you, one of the secrets of AI is the more power and tools that you give these models, the better they get. It seems really obvious, but I've seen a lot of engineers—

that's why you want to connect everything.

That's right, give them access to email, they give them access to Calendar. But you're going to need technologies like a firewall, you're going to need virtualization so that, you know, if, if it goes wrong, you can shut it down or you can snapshot it and restore it. So a lot of these workloads are gonna move to the cloud. So there's reason to be bullish on cloud infrastructure and the large platforms that are gonna run millions or trillions of these digital workers, which is how I like to think about them.

So, okay, so one guy created ClaudeBot, which became MultBot, which is now OpenClaw. One guy created that, and this was basically enabled by what Claude Code and Opus 4.5 was able to do. He said it was very, very sophisticated under the hood. My understanding, correct me if I'm wrong, is that Anthropic kind of put the brakes on it. And so people started using a Chinese open source model, Moonshot's Kimi, to do all the work in the background. And it's cheaper and more efficient. And I just saw a stat that Kimi is now— this is amazing, top of the leaderboard in terms of token usage on OpenRouter. How much of that is because of OpenClaw and was it wise for Anthropic to put the brakes on it?

Look, one of the challenges with AI, and I think the reason for us to be excited is the demand for it is like nothing we've ever seen in the history of technology. Vercel operates a global token delivery network called the AI Gateway. And what we let developers and companies do is very securely obtain tokens by vetted sources, Anthropic, NVIDIA, different inference partners and providers, OpenAI, of course, Gemini, et cetera. And we're also seeing a lot of change—

Like the Chinese models.

Mobility, exactly. So we're seeing a lot of change, for example—

You do include the Chinese models though from your offerings?

Yes, from vetted providers that, you know, have the weights and have security built in, et cetera. But developers run into rate limits with providers or they run into price performance issues with the large AI labs. And so the open models are clearly offering an attractive option. I was talking to a Vercel customer recently who moved from GPT Nano to a Qwen fine-tune, which is a Chinese model, saved a ton of money and improved performance. So of course open source is going to continue to make a lot of progress. Now, I do— I don't think they're anywhere near the frontier. They get close each time and they, and they continue to sort of bridge the gap. But it looked like KIMI was getting close. And now today we have Opus 4.6 and Codex 5.3.

And tomorrow we'll have Gemini something.

America's winning. You know, the president was sharing this a couple of days ago. I think we're so ahead in AI. I want to make sure that we keep that lead and I want to make sure that developers keep getting awesome open source options.

Okay, that's developers. But I mean, most people don't really care that they're getting the frontier, right? I can imagine my brother's always asking me, you know, what he— what chatbot he should be using to make images when he runs out of his free tier on ChatGPT or Claude. And I told him, go to the Chinese models, right? Because they can do it sort of just as well for his purposes. If we live in a world where everyone has an Open Claw, an AI agent that acts autonomously, do they need Opus 4.6 or the latest GPT to power it? Why wouldn't it be an open source model if they're cheaper and more efficient and almost as good?

Yeah, I think when the agent is doing a very broad set of tasks, because the interesting thing about what we're seeing with agents is that, and I think you were calling out, I shared that 87% of Vercel support tickets are now being automated by, by our own agent that we built and deployed on Vercel. 87% of some of the most difficult support tickets in the world, because we have so many sophisticated technical customers, as they're being beautifully automated by agents. And, and what we had to deploy there is not a Chinese model. I needed the absolute most intelligent model. Who would want to not get the best possible support experience, right? When you're stressed, when you're maybe you're just considering switching to a different provider, competitor, et cetera. So I do believe that when it's knowledge work is involved, very, very capable, generally effective models are needed. Now, sometimes we might need—

I understand that argument, but the consumer market is just massive, right? And if you really believe that everyone is going to have these tools, have an agent, right? Isn't that a bigger market?

Well, well, where I see the difference is we saw a lot of great announcements from OpenAI about solving frontier science problems, solving Erdős problems in mathematics. I'm very excited about the prospects of mathematical superintelligence, breakthroughs in physics, medical research. All of these, when you look at the benchmarks as well, we're spending huge amounts of time and tokens. And so there's definitely a trade-off between how fast the model responds and the capabilities of its reasoning, right? So you don't wanna throw a gazillion tokens at a simple problem saying like, "Hey model, how are you?" "Good." Like you don't need the best model possible for that. And so technologies like Vercel AI Gateway over long periods of time are gonna enable you to automatically route across a huge portfolio of models. So if you greeted the model and say, "Hey model, thank you." We'll use the cheapest model available. But if you give me like a mathematical breakthrough task, that's right. Model agnosticism is the future.

But where, like, okay, so what are the majority of tasks right now? What are going to be the majority of the tasks that will need a cheaper, not a frontier model in the future? Like, what's the breakdown right now? Where do you see that going?

So right now, what I'm seeing from Silicon Valley is Everyone wants to create agents. Everyone wants them to be as intelligent as possible. Models like Opus 4.6 are clearly the— To link it back to MoldBot, what's fascinating about MoldBot is under the hood, it uses the technology that backs Claude code. The coding model that the engineer uses is Codex. He's dating both Anthropic and OpenAI to create this project because they're using the strength of each. And so I do encourage developers to not lock themselves in to a particular stack. And things like Vercel AI Gateway basically allow you to effectively do that. The fluidity you mentioned, I think Google is going to make a big comeback as well. I mean, comeback because they were hot 3 weeks ago, right? So to speak in relative terms. But in xAI as well, I think everyone's realizing that going back and refocusing on coding, kudos to Anthropic for focusing. To get that advantage gives great results.

So the lead, you have the lead for such a short amount of time, and then we've got another model coming out from Meta, right, which is apparently the best that they've ever seen internally. If I had to, like, if you had to choose right now, Guillermo, which company is winning, who is that?

So again, very dependent on each task. I'll give kudos to Google.

You gotta have an overall. No hedging. Okay, fine. You can hedge a little bit. First, if you had to choose, like gun to your head, had to choose one model, which is, and then I'll let you nuance.

Okay, I'll tell you, uh, a baseline agent infrastructure tool calling, obviously Anthropic. Uh, image, Nanobanana by Google. Um, video, I think Google also has, has a relatively smaller perhaps lead, but definitely, uh, Veo is doing a great job. And I'll say, I think on really hard research tasks and long-running reasoning, like you need the absolute best intelligence available, GPT-5 Pro has an insurmountable advantage, it seems.

Does that show up in the benchmarks, what you're saying about ChatGPT? Because I feel like what I've been looking at recently, it isn't.

I advise caution when looking at benchmarks. A lot of them are gamified. A lot of benchmarks don't tell the full story. It's fascinating to see how you mentioned the vibes, the press. Even though we're at the cusp of singularity and superintelligence, a lot of us are still navigating the world by what we perceive to be effective. A lot of our product development methodology at Vercel is we test things out with customers. We have to launch experiments. Our support agent was rolled out progressively to ascertain that people were enjoying the customer support interaction. So, so much of this is still to develop, to deliver a great customer experience and ultimately, you know, mitigate and minimize the slop, which is sort of the antithesis of AI. I don't want to live in a world where we're just, we take a benchmark at face value and then we ship subpar products to customers.

And that's why I ask you that question, because that's great to know. And I think like also for just average users out there, they've had ChatGPT and that's been the go-to, but people are starting to wonder what else. Do you think— I mean, I don't know if you saw the latest stats, Gemini. I mean, Sundar Pichai said this on the call last night. What was it? Over 700 million users. That's closing the gap pretty quickly. Is this still anyone's race?

Well, I mean, Google is pushing their AI products really hard. Every time I use a Google product, I get 3 popups to use some of their AI. So there, there's also question about, you, you know, is this a, is this a user that's enjoying the AI product? Is this a user that just can't wait to uninstall it because it's being shoved down their throats? Uh, so I'll tell you, uh, companies are right now in a race to claim that all of their revenue is AI revenue, right? Uh, and so at the end, at the end of the day, you have to look at stickiness. You have to look at retention. You have to look at NPS. Uh, and, and that's the ultimate measuring stick of these AI capabilities.

I do wonder about the cost of all of this. I mean, when I try to run a prototype on Cloud Cowork or Replit or anything really, I can just kind of see in real time like the number signs, you know, just adding up in terms of the amount of compute it takes. There's no way, right, that like a Max Pro subscription pays for everything that everyone's doing, especially like kind of dumb stuff that I'm trying to make?

Well, I'll tell you, each of those prompts could be saving you weeks of time. When you do something with Cloud Code, in the past, we were orchestrating teams of people to do the same task. Design, engineering, product management, conversations, meetings, et cetera. Now it's a prompt. And what's really exciting is these models are getting really good at deploying these applications to production. A couple days ago, we announced our skill built into Codex for Desktop, where you prompt— you should try this at home— tell Codex, hey, build me this idea, and then deploy to the cloud. And it's going to recommend Vercel to host and scale that application for you. And like I mentioned earlier, this is the same infrastructure that scales the largest websites on the internet. It's an, it's an incredible time to be alive where the average human being can now participate in this global infrastructure that before was only available to FAANG. It was only available to the 6 largest companies on the internet. Those were the only ones that could actually scale and do horizontal scalability and a lot of these tech feats.

You've inspired me to come to you now. Jasmine and I will come to you when we set up our OpenClaw intern. You made that feel a lot more possible. I think security is, you know, top of mind for myself. For a lot of people. Guillermo, we'll wrap up in a moment. Let me just ask you, because we started this conversation saying that the last few weeks have just been like just so wild, it's really hard to explain it to people outside of tech or outside of Silicon Valley. Just these huge, enormous leaps that have been made in AI and even the things that someone like myself, a non-technical user, is doing. What are we going to say a few months from now or a year from now? What do you think the biggest change is going to be?

So right after New Year, I had a meeting with my entire company. I told them, look, this is yet another moment where there's been a discontinuity in the growth of capabilities of AI. And I encouraged everyone to think about— think back to the things that you used to say AI cannot do. Oh, AI cannot do large projects. AI cannot do front end and back end. AI cannot do scale. Every single thing we've said AI cannot do gets proven wrong 2 weeks later. And so, I always tell people, get really good at using these AI tools. Your leverage will be immense. One person, billion-dollar company, well within sight. Autonomous companies, well within sight. Completely self-driving infrastructure like Vercel where You don't have to configure anything about the cloud in order to scale to, you know, travel the entire world with your ideas within Sight or even Reel today. So get really good at using these agents, build your own. I think this is the next frontier of scale. Like the individual can now provision their own agents. You know, I try to download a lot of my insights into my own agents so I can scale myself.. And so it is a time where I think the individual is actually ironically really empowered.

Right. Maybe protect your data too, right? So, so no one else can scale yourself on behalf of you. It's amazing. Just to that point, I was reading that vibe coding only became a term coined by Karthi like a year ago. And I do remember trying to do this and I couldn't do it until a few weeks ago. I thought it's not for me. Clearly you need to be some kind of developer, have you had to have some kind of experience with code. And now you just don't. And I encourage people who are watching this to go do it. Jasmine and I can't stop. We're just constantly all day, every day, how can we make this CNBC segment? How can we test it? Like, how can we actually build something to show our audience? And it surprises us, you know, day after day. Guillermo Rauch, this was a wonderful conversation. Come back and talk to us again. There's no shortage of news ever. The model, the iterations, everything just keeps coming so quickly. So we hope you'll come back, and thank you for taking the time.

See you next time.

And a special thank you to producer Jasmine, who's doing all of this with me, Janice, who's behind the camera and, you know, engineering all of these live streams, and of course Robert and Evan on the technical side. Thank you guys for watching. We'll see you again soon.