Photo-Illustration: Intelligencer; Photo: Getty Images

A little less than a year ago, the AI industry acknowledged wider anxieties about a possible bubble. Sometimes, OpenAI’s Sam Altman said, “smart people get overexcited about a kernel of truth.” Mark Zuckerberg allowed that a bubble is “definitely a possibility, at least empirically,” while Google’s Demis Hassabis suggested that “some parts of the AI industry are probably in a bubble.” Jeff Bezos, while arguing that AI will “change every industry,” pointed to signs of an “industrial bubble.”

A few months later, this bubble-talk bubble was pricked by the arrival of a new generation of AI coding tools, which were notably more useful for complex, real-world programming and quickly adopted by the tech industry. This had a few different consequences for AI firms. It refreshed the industry’s narrative of inexorable progress, even if they hadn’t quite anticipated how things would play out: The lines were still going up, and insiders once again described feelings of acceleration. One month, prominent AI figures were conceding that most AI-generated code was “slop,” and the next, they were delegating a significant portion of their work to Claude Code. In addition to a narrative reset, AI firms now had something to sell to other firms, which showed immediate interest — a plausible business model beyond getting as many users as possible. Lastly, frontier AI coding tools were extremely compute-intensive. They could write code but required enormous quantities of tokens to do so. This resulted in some sticker shock among AI customers, as well as some unsustainable subsidies on the provider side, but also quelled, at least temporarily, concerns about overbuilding data centers. For the first half of 2026, capacity was scarce, major AI firms were selling as much access as they could provide, and plans for hundreds of billions of dollars of infrastructure buildout regained credibility among investors. Problem solved!

For a while. The anxious mood of late 2025 was caused, in part, by worries that model progress was at risk of slowing down. The arrival of agentic AI coding relieved that. But another contributor was the arrival of Chinese open-weight models, which were cheap to use, free to run on customers’ own hardware, and anywhere from three to nine months behind the American frontier. In mid-2025, a model called DeepSeek R1 promised similar capabilities to OpenAI’s GPT-4, accompanied by claims that it had been trained for a fraction of the price, briefly crashing American tech stocks. Now, in the post-AI-coding era, Chinese AI firms — which American companies have since accused of “distilling,” or ripping off, their frontier models — are once again entering the picture. About six months ago, Anthropic and OpenAI passed a new threshold in AI software development. Now, multiple open-weight competitors, according to AI researcher Nathan Lambert, are passing it too. These include Moonshot AI, which makes a model family called Kimi, and Z.ai, which just released an update to its GLM model:

Pretty much everyone I respect among the AI commentariat and researcher class has praised the model after using it personally. Such a focal point of discussion among the community has only been so clear with an open model release once before — DeepSeek R1. This is not a comparison I make lightly, and when I compared Kimi K2’s release to a “DeepSeek Moment,” GLM-5.2 has well exceeded that… 

GLM-5.2, he writes, is the first open-weight model “that feels right in coding harnesses as a general agent,” which is to say it’s the first of these vastly more affordable AI models to be able to do most of what people expect from using Claude Code with a recent Anthropic model, or the equivalent from OpenAI. For the moment, it appears as though these models are at least as good as anything Google has on offer, if not better, which is an outcome that would have sounded strange just a few months ago. (For its part, newer DeepSeek models are starting to show up in corporate AI spending.)

The broader consequences of commoditizing circa-early-2026 agentic coding aren’t easy to predict. For one, the frontier labs have kept moving. Since then, Anthropic released (and then was forced by the government to recall) Mythos and Fable, a new generation of models from which another set of new capabilities, this time around cybersecurity, emerged unexpectedly, meaning that frontier models still had something unique, valuable, and expensive to offer to customers. (As of this week, however, OpenAI claims to be able to offer similar capabilities at a lower price just weeks after Anthropic’s release; Z.ai claims its own models will be there by the end of the year.) The arrival of cheaper, good-enough coding models could undercut offerings from American firms, which have raised many times more money than their Chinese counterparts. Alternatively, these cheaper models could end up inducing demand, making heavy coding-agent usage more attractive to a wider group of customers without cutting too much into the top end of the market, where the frontier labs could still dominate.

Then there’s the vaguer prospect of another breakthrough that might render six-month-old capabilities less relevant to the bigger picture: AI companies are talking an awful lot about “loops” these days and hinting at the possibility that their models are getting closer to “recursive self-improvement,” which they argue could solidify their leads in profound ways. It would be wrong to say that stranger things have happened — a self-improving AI that threatens to slip from human control is about as strange as things get — but things that are at least pretty strange have, and it would be imprudent to dismiss the possibility that, given the billions of dollars of research money and compute thrown at the problem, the next generation of models might be useful for something else the economy values as much as coding. At the very least, the industry is looking for its next big story.

But if you zoom out a bit, the basic dynamic here is punishing and carries undeniable risks for the industry. With each step-change in AI capabilities, affordable, flexible, and open models are pretty close behind, and what was briefly expensive to deploy becomes quite cheap. American AI firms are openly worried about this sort of thing, which they’ve framed as a threat to national security, a form of intellectual property theft, and a source of uncontrolled AI risks, but which also clearly complicates their still-nascent business models. Bubble talk isn’t quite back, but there are early signs — among tech leaders, but also in markets, which are suddenly cooling on AI and chip stocks — that vibes may be shifting once again, as they have a half-dozen times in the last three years.

Sign Up for John Herrman column alerts

Get an email alert as soon as a new article publishes.

Vox Media, LLC Terms and Privacy Notice