中国开源AI模型GLM 5.2逼近前沿，成本仅五分之一

Deepsec showed Wall Street that China could build cheap, powerful AI. Now, Jeep the AI is showing how that advantage could spread. >> I've been consistently surprised by how quickly the open source has caught up. And I think GLM 5.2, you're kind of seeing the first model where it's really competitive with like Opus 4.7 and some of these like very closed source frontier models. What JLM52 kind of shows is, you know, we might still be in this sort of three to six month territory. Everything about the economics of AI changes depending on which outcome we're in. >> We need cost to come down. That's how you drive adoption. That's how you get return. >> The AI leaderboard is starting to look incomplete. The next fight may be over intelligence per dollar. Models that are capable enough for real work and cheap enough to run constantly. That is where Zai is forcing a new conversation. Japu's latest model, it's called GLM 5.2, comes out of China. It has landed with a bang in Silicon Valley, blowing past all other open- source models, nearly matching the American frontier for just a fraction of the price. Developers, they're piling in. Open router token traffic showing much quicker adoption for GLM52 than DeepSk's V4 launch back in April. It was a big deal then, it's a big deal now. This just carries this story that we've been reporting on even further. It all matters because GU is hitting at a different moment though. The trillion dollar sell-off post deepseeek that was treated as a kind of a one-off shock partly because people saw it as a chatbot story. GLM 5.2 this is different. It's strong at agentic work and that is key on one agentic benchmark. It is just one percentage point away from opus 4.8 for a fifth of the cost. Now opus 4.8 that is anthropic most powerfully available model. So basically you're getting very close to the same horsepower for just a fraction of a cost. And it really comes at a time when expensive AI is already eating into budgets. And that is a really hard trade-off for enterprises, for companies, for Fortune 500 to ignore. Aentic AI, it's only going to intensify those costs. More complex tasks, more steps, more tokens. So now that GU or also known as Z.AI is in the picture, a Gentic open- source, it might be the next big threat out of China. And here's the part of the story, the AI story that I think Wall Street is still missing here. For the last few years, everyone has been obsessed with these AI leaderboards. Who has the smartest model?

Who's the best at coding? Which one is ahead on reasoning? But that is not how companies buy software. Companies are increasingly asking what is good enough and what does it cost to run this thing a million times or more. All of my employees running it a million times over the course of weeks, months, a year. So the new metric in AI, it is intelligence per dollar. And one reason Chinese models are pushing so hard on that metric is because of distillation which is basically you take a big expensive model use it to train a smaller cheaper model to act like it. The American AI story it has been built around bigger models bigger data centers huge spending. But Chinese labs they're putting these cheaper versions front and center very very close to the frontier performance without the costs. That's why GPU is interesting. Artificial analysis has this chart that looks at both sides of this. How smart is a model and how much does it cost to run it?

Now, the most attractive quadrant, it's obvious. High performance, low cost. That's where every company wants to be. Jeep's GLM 5.2. It's getting very close to that sweet spot. It's not quite at the top in terms of performance. You can see it here, very close to that green square. So, not at the very top in terms of performance, but it is close enough to the best models from OpenAI and Anthropic to make this price gap hard to ignore. And on agentic work, that gap matters even more. These are tasks where models don't just answer one question. They plan, they code, test, fix mistakes, they loop, they keep going, and that is a lot more expensive. So, if you're a company, you need a great model, cheap enough to use over and over again. And that is the setup. I want to get into all of this with two people who are actually living this. Aaron Levy, he's been running Bucks for 20 years. He saw the move to the cloud, the move to mobile, the move to AI. He is one of the most plugged in enterprise guys in the valley on what this stuff actually looks like when it hits a real company. We've also got Gabe Pereira, co-founder of Harvey, which is the AI platform um that half of the AM law 100 is now running on. Now, before that, he was a research scientist at DeepMind and Meta. So, he has seen this from both sides, the model layer and the application layer. Erin and Gabe, it's great to have you both on. We got Aaron, too. Okay, we're gonna I can't Let's see. Do we have both their sound?

>> Okay, we're going to try and get this sound fixed. So, Erin, you're you're you're an observer now. I'm assuming that you can hear Gabe and I, and then I'll come to you on this. Um, Gabe, let me start this off with what's happening with, you know, GLM. It feels like my entire X feed. Everyone I'm talking to here just cannot get enough of this new model GLM52 right now. Um kind of feels like deepseek all over again. People also aren't being shy in calling that maybe even bigger. So I guess first how do you explain this to someone outside of tech or like a non- tech CEO um who has been allin on American AI but sees all of this buzz around a new model that is almost as capable but a lot cheaper. >> Yeah. I I would say the big question since kind of chat GBT came out and these closed source models is just how big is the gap between the closed source and the open source. And I think I've been consistently surprised by how quickly the open source has caught up. And I think GLM 5.2 you're kind of seeing the first model where it's really competitive with like Opus 4.7 and some of these like very closed source frontier models. And exactly to your point, it lets you think about the cost curves, where you need frontier intelligence, where you can use open source intelligence. And I think this is what we're starting to figure out and a lot of companies are starting to figure out now, >> right?

So explain that kind of in like just basic terms. If you are a company CEO and you have this AI budget, but you've seen, you know, your company blowing through it, when you get something like this, what should they be thinking?

Especially when you say like it almost reaches the frontier, the very best models but for a fraction of the price. People you are talking to, Gabe, do you think that they're taking a second look at Chinese and these open source models or should they still be concerned about them?

>> Yeah, I think every company will use a mix of both. But exactly to what you said of starting to think about which tasks do I not need to use frontier intelligence?

And so as these models get smarter, do you need the smartest model reviewing your contracts or doing simple red lines? And so I think for every company, the same way you organize a large company by figuring out, you know, I need different seniority, I need to pay different people for different roles. I think you're going to start seeing this with agents where there'll be some agents that are making critical decisions. You want frontier intelligence and there'll be some agents that are doing simpler tasks and there you can use open source and you can really reduce the cost of of intelligence. >> Right. And are you seeing sort of the buzz that I am as well?

is your feed, the people you're talking to, are they very excited about GLM52 and like can you explain why this is so interesting, why this is different than Deepseek?

I tried to, but I'd love to hear it from you. >> Yeah, I think we're seeing the same buzz. It in terms of our benchmarks, it was kind