Why Cheap Chinese Open Models Are Taking Over the AI Market in 2026

Chinese open weight models went from almost nothing to the majority of usage on the largest neutral AI marketplace in eighteen months. Here is what the data actually shows, why it is happening, and the important part the headlines leave out.

• Here is a number that sums up the biggest shift in Artificial Intelligence (AI) this year. In June 2025, AI models made by American companies (OpenAI, Anthropic, and Google) handled around 70 percent of the traffic on OpenRouter, the largest neutral marketplace for AI models. One year later, in June 2026, that share had fallen to roughly 30 percent. Models made in China now handle most of the rest.
• That is a huge change in just twelve months, and it has set off a lot of dramatic headlines about American AI collapsing. The real story is more interesting than the headlines, and more useful if you are trying to decide which tools to actually use. Let us look at what the data shows, why it is happening, and the important part most coverage leaves out.

First, What Is OpenRouter and Why Does It Matter?

OpenRouter is a marketplace that lets developers send their work to hundreds of different AI models through a single connection. Instead of signing up separately with each AI company, you go through OpenRouter and it routes your request to whichever model you choose.

This makes it one of the cleanest signals we have for what developers actually use, because it is neutral. Nobody is locked into one brand. When a model wins traffic here, it wins because developers chose it on merit and price, not because of marketing. OpenRouter now processes more than 20 trillion tokens every week. A token is roughly three quarters of a word, so that is an enormous amount of real work flowing through the platform.

Keep that word neutral in mind, because it becomes important later.

The Numbers: A Fast and Real Shift

The change is not gradual. In late 2024, Chinese models were barely a rounding error on OpenRouter, under 2 percent of usage. The turning point came in the week of February 9 to 15, 2026, when Chinese models processed more tokens than American ones for the first time ever on the platform.

Since then the gap has widened. By mid 2026, Chinese open weight models account for somewhere between roughly 45 and 60 percent of usage depending on how you count, and they hold several of the top spots on the leaderboard. DeepSeek, Alibaba's Qwen, Moonshot's Kimi, MiniMax, and Z.ai's GLM are the names doing it. American models did not get worse. The market simply expanded enormously, and most of the new volume went to the cheaper option.

Source: OpenRouter usage data reported by Bloomberg and Exponential View, June 2026.

Why Is This Happening? Price, and an Ironic Twist

The single biggest reason is price. Chinese open weight models are often 10 to 35 times cheaper per token than the leading American ones. When you are building a product that processes millions of tokens, that difference decides your bill.

Here is the ironic twist. For two years the United States tightened export controls to stop advanced AI chips reaching China, expecting it to slow Chinese labs down. Instead it pushed them to get more out of less. They leaned hard into efficient designs called Mixture of Experts, which only switch on a small part of the model for each request, cutting the computing cost dramatically. The restriction meant to widen the gap helped close it.

The other driver is what people use AI for. Writing code grew from about 11 percent of OpenRouter usage at the start of 2025 to more than half by mid 2026, and Chinese models are both strong and cheap at coding. As coding agents took over, the cheap models rode that wave.

The Part the Headlines Leave Out

Now for the context that turns scary headlines into something more accurate.

OpenRouter is not the whole market. It is the place developers go when price is the main thing they care about and they have no reason to stay loyal to a brand. But anyone using a top American model usually goes straight to that company's own service, because there is no cost saving in routing it through a middleman. So OpenRouter naturally undercounts the American leaders. It is a leading signal for the price sensitive layer, not a full census of all AI use.

This is why two separate markets are forming side by side. One is a commodity race where the cheapest good enough model wins, and Chinese labs dominate it. The other is a premium tier where businesses pay more for top quality, safety, and very long memory, and the American labs lead it. The clearest proof: on the same platform, Anthropic held about 12 percent of token volume but earned about 46 percent of the revenue. Losing cheap volume while keeping expensive revenue is not collapse. It is the market splitting in two.

For scale, Anthropic reached around 47 billion dollars in annual revenue run rate in May 2026. A company sliding toward bankruptcy does not post numbers like that. The honest read is that American labs are losing the race to be cheapest while still winning the race to be most valuable.

Source: OpenRouter revenue and volume figures and Anthropic revenue reporting, mid 2026.

The Export Control Angle

There is a real policy story here too, though it is more measured than some versions you may have seen.

In June 2026, a United States export control directive required Anthropic to switch off broad access to two of its newest models, Fable 5 and Mythos 5, in order to prevent access by people outside approved groups. Just days later, on June 13, the Chinese lab Z.ai released GLM 5.2, a powerful model anyone in the world can download and run with no restrictions. The timing was not lost on anyone. When a top closed model becomes harder to reach, an open one that is free to download becomes more attractive, especially for organisations that need to know their tools will not be switched off.

It is worth being careful with the framing here. This was an export control rule about who can access certain models, not a blanket ban on American AI, and the situation continues to evolve. We mention it because it genuinely affects which tools are available to whom, which is what our readers care about.

What This Means If You Are Choosing Tools

Strip away the geopolitics and here is the practical takeaway.

If cost is your main concern and you are building something that processes a lot of text, the cheap open models from China are now genuinely capable and worth testing. The price gap is real and large.

But price is not the only cost. If you send data through a Chinese model's own first party service, check the terms carefully. Some allow the provider to train on what you send, and some route your data through servers in China. For sensitive work this matters a great deal. The good news is that these open models can also be run through Western hosting providers that promise not to train on your data, often at similar prices, which removes much of the concern.

The simple rule: pick the cheapest model that meets your quality and privacy needs, not just the cheapest model. For a lot of everyday work the budget option is now good enough. For sensitive or top quality work, the premium American models still earn their price.

The Bottom Line

The story is not American AI collapsing. It is the market growing up and splitting into two: a cheap, fast, open layer where China now leads, and a premium layer where the American labs still command the highest prices and the most revenue. Both are real, and both matter depending on what you are building.

At AI Tools Mentor we track the prices of all the major models, American and Chinese, and re verify them every Tuesday and Friday, so you can compare the real cost of each option for your own work rather than guessing from headlines.

Sources: OpenRouter usage and revenue data, Bloomberg and Exponential View, Artificial Analysis, and provider reporting, all from mid 2026.