3 times cheaper and no worse than ChatGPT: OpenAI's nightmare comes from China
The open-source model Kimi K3 from China is beating performance benchmarks, writing code, and threatening the American business model. Despite sanctions and the pursuit of distillation in Washington, the Chinese strategy of cheap, high-power models already accounts for a third of traffic in the US. Now, it is igniting a regulatory and economic war over the future of AI.
It is happening again. Just when it seemed that American AI labs had reopened a gap with the launch of new models from OpenAI, Anthropic, Meta, and xAI, a new model arrives from China that undermines this feeling. A few days ago, Moonshot AI unveiled Kimi K3 — an open-source language model that immediately ranked at the top of performance benchmarks and reignited a much broader question than the quality of any specific model. Is the Chinese strategy of open and cheap models starting to threaten the business model upon which the American AI industry was built?
Moonshot demonstrated Kimi's capabilities through a series of demos. An AI agent that developed a computer game, another agent that wrote a compiler for programming graphics processors, and even an agent that autonomously designed a small AI accelerator within 48 hours using open-source chip design tools. The interest was so high that within two days, the company announced it was temporarily freezing new registrations after demand exceeded its computing capacity.
Kimi K3 includes about 2.8 trillion parameters, making it one of the largest open-source models in the world. Despite its size, it is based on a Mixture of Experts architecture, where only a small part of the sub-models is activated for each request, thus achieving high performance while using computing resources more efficiently.
"What makes this achievement particularly impressive is the way the model was built," says Shai Friedman, CTO at the software house CodeValue. "US export restrictions limit China's access to the most advanced chips, so instead of relying on more and more computing power, Moonshot's engineers managed to get more out of limited resources. This is an architecture that puts efficiency at the center and was born out of constraints."
Price is also part of the story. Moonshot charges about 3 dollars per million input tokens and 15 dollars per million output tokens, compared to about 5 and 30 dollars for GPT-5.5 and 5 and 25 dollars for Claude Opus 4.8 — a gap that could save millions of dollars for organizations running large volumes of queries.
The battle is no longer just about the best model
For the past two years, OpenAI and Anthropic have tried to establish an advantage through closed models and massive investments in computing. Conversely, Chinese companies like DeepSeek, Qwen, and now Moonshot have chosen a different strategy: open models, low prices, and wide distribution.
This strategy is already being felt on the ground. According to data from Vercel's AI Gateway, used by companies like Airbnb, Shopify, DoorDash, and Cursor to route requests between AI models, Chinese models like Qwen, DeepSeek, and Kimi are already handling nearly a third of the traffic — three times their share from three months ago — even though they account for only a small fraction of the total cost. "This is not ideology, but simple economics," says Friedman. "Every task is routed to the cheapest model that provides a good enough result — and more and more often, that model is Chinese."
The tension is not just about price, but also about intellectual property. OpenAI and Anthropic claim that Chinese companies use distillation — a method where a new model learns from answers generated by another model — to accelerate the development of their models. In the case of Kimi K3, no public proof of this has been presented, and many believe that its capabilities also reflect genuine engineering innovation that is difficult to explain by distillation alone.
The front moves to Washington
The concern in the United States already goes beyond technological competition. Investment manager Ruchir Sharma defined the American economy as "one big bet on AI." According to the data, since the launch of ChatGPT, AI-related stocks have been responsible for about 75% of the S&P 500 index's return, about 80% of earnings growth, and about 90% of the increase in capital expenditures by companies in the index.
The Chinese strategy threatens exactly this assumption. As China did with electric vehicles and solar panels, it does not have to lead at the high end of the market. It is enough that it offers cheap and good enough alternatives to pressure prices and erode the profitability of competitors.
In Washington, there is already a debate about whether to restrict the penetration of Chinese models. Investigations have been opened in Congress regarding the use of Chinese models by American companies, and at the same time, a bill has been introduced that would prohibit federal agencies from using them. Dicken Bull, head of strategy at OpenAI and a former AI advisor in the Trump administration, estimated that the administration might create "large amounts of regulatory risk" around Chinese models, so that companies would stay away from them even without an official ban.
On the other hand, David Sacks, the White House AI advisor, argued that OpenAI and Anthropic already hold a "duopoly" on AI model revenues and are trying to use the government to block open competition. According to him, regulation that would delay data centers and model development in the United States will only help China. "That is how you lose the AI race," he wrote.
The security concern, however, is not invented. Banks, health companies, government bodies, and security organizations are not rushing to entrust sensitive information to Chinese suppliers, so American labs still enjoy an advantage of trust, legal liability, and control over the supply chain. If DeepSeek was the surprise that proved China is capable of closing the gap, Kimi K3 already looks like a trend. The question is no longer whether China is capable of building models that compete with the best American labs, but whether the strategy of open and cheap models will also change the economic rules of the AI industry — and how the United States will choose to respond.
