American AI coding agents are impressive. But so are China’s

There’s a new Chinese model gaining buzz among developers: Zhipu AI’s GLM 4.7. That in itself is not new, but what did catch our attention was where it’s becoming popular: the U.S., our own turf.

I first spotted Zhipu’s post on WeChat, announcing that the new coding tool was seeing so much demand, it would start limiting access. A year after DeepSeek’s R1 model shook the artificial intelligence landscape in the U.S., AI experts are saying that they’re seeing Chinese AI models spreading around the world. I wanted to see if that was also true for Zhipu.

I reached out to the company, with no expectations of hearing back, just as DeepSeek has never responded to our requests. But Zhipu recently went public in Hong Kong, and its investor relations team is on top of the ball. They replied almost immediately, telling me that “the user base of Zhipu GLM Coding Plan is primarily concentrated in the United States and China.”

American developers have always told us of the bias against using Chinese models — so for Zhipu to be gaining traction here suggested a real DeepSeek-like breakout moment.

Just last week, we were floored by the apps that Replit and Claude Code were able to produce in minutes. They felt like the frontier of AI, an example of how American innovation is leading the way. It’s also creating a tangible shift, leading to a 60% surge in new app releases. But if a model out of China is just as powerful and easy to use, are they really six months behind like Google DeepMind’s Demis Hassabis says? And if they’re cheaper and open source, what real moats do the American AI coding agents have? 

CNBC’s Deirdre Bosa and I jumped on the chance to experiment with Zhipu, to see if the hype would match our successes with Replit and Claude. I’ve always found Chinese stock exchanges difficult to navigate and get data from, so we asked it to build a tracker for the biggest public companies in China. Zhipu whipped up an app more quickly than its American counterparts could, but its results were less polished.

We’re naturally skeptical of claims from companies that tout usage without publishing metrics, so we checked in with some sources — people building in AI that would come across Zhipu if it really was making inroads here. It was quickly clear that Zhipu’s GLM 4.7 is indeed a model gaining recognition in the U.S., though reviews are mixed.

One of the builders testing it is Tuhin Srivastava of Baseten, a platform that sits underneath AI applications and sees real-world model usage across enterprises. That matters even more now that Baseten just raised a new round with Nvidia participating, a signal that it’s plugged into real, large-scale AI workloads. 

A conversation with Srivastava and Bosa was livestreamed here at 2:30 p.m. ET on Monday.