The hottest AI models in Silicon Valley face a powerful source of competition
Burdened by rising costs from labs like OpenAI and Anthropic, U.S. companies have begun to shift to Chinese models in a landmark adoption of the country’s software.
As Anthropic and OpenAI rocketed to popularity based off of closely guarded artificial intelligence models, Chinese AI companies have made a different bet: They would make theirs open and free.
Over the past year, the strategy has grown more influential, especially in the wake of the U.S. government’s now-rescinded restrictions on flagship Anthropic model Fable. Burdened by rising AI costs, American tech companies have begun to shift to Chinese model families such as Alibaba’s Qwen, Z.ai’s GLM and Moonshot AI’s Kimi — a landmark adoption of Chinese software at enterprise scale in the United States.
“This is the first time that it’s happening,” said Rafiq Dossani, an economist at Rand. “China’s been able to catch up in software at a rate that it couldn’t do much earlier, even though [it’s] well resourced.”
ChatGPT, Claude and Google’s Gemini are all examples of “closed” AI models: The secret sauce of how they produce text is tightly guarded within their companies, and the only way to access them on an enterprise scale is to buy a subscription or use costlier pay-as-you-go pricing.
In contrast, many leading Chinese models are open-weight, meaning that their internal formulas can be inspected and accessed for free. While the companies also offer their own web interfaces and desktop apps, most outside developers choose to just download a copy of the model and, with some configuring, hook it up to their systems as they see fit. That means the main cost of using open-weight models is the computing power needed to generate text — generally far cheaper than a closed model in a big lab, where training and development costs are priced in.
The savings from open-weight models are especially stark following the trend of “tokenmaxxing,” when some companies in Silicon Valley urged their engineers to pump up their AI usage as much as possible as a sign of productivity. Tokens are the basic units of input and output in language models.
Non-Chinese companies, including Meta, Google and France’s Mistral, offer open-weight AI, too. But in the global race for AI dominance, the most popular open options come from China-based companies, which began outpacing American counterparts in downloads in August.
On OpenRouter, a popular marketplace to access and run open models, Chinese AI accounted for nearly half of U.S. traffic by tokens in the last week of June, up from 16 percent at the beginning of the year.
Greg Osuri, CEO of the cloud computing company Akash, spent almost $12,000 on Claude by himself in May. But after the U.S. government instituted export controls on Fable, Anthropic’s flagship model, he started using Z.ai’s GLM-5.2, which he said was cheaper and better than Anthropic’s models.
“I switched to GLM-5.2 because I’m not paying top dollar for a model that’s inferior,” he said. “And I don’t see a reason why I should switch back.”
Osuri has more than halved his personal AI spending since. On his network, 1 million tokens of output from GLM-5.2 — about 750,000 words — costs $4.40. The same workload on Claude Opus 4.8 is $25 when using Anthropic’s pay-as-you-go pricing.
In addition to being used in internal coding tools, Chinese open-weight models have been used to power customer-facing agents. Major cloud computing providers, including Cloudflare, Amazon Bedrock and Microsoft Azure, are offering access to Chinese open-weight models on their networks (Amazon Executive Chairman Jeff Bezos owns The Washington Post). Ottawa-based Shopify has experimented with using Alibaba’s Qwen to power an AI assistant for vendors on the e-commerce platform. The company said in a blog post in April that the open-weight model was 68 percent cheaper than a closed option. Shopify did not respond to multiple requests for comment.
Chinese open-weight models first burst onto the scene in early 2025, when DeepSeek released an open-weight model with similar capabilities to those of the major U.S. labs, but at a fraction of the training and deployment cost. The most recent leader has been the Beijing-based Z.ai, whose GLM-5.2 model has posted performance comparable to everyday models like ChatGPT 5.5 and Claude Opus 4.8 on coding benchmarks (although more heavyweight versions of American AI post better performance). In turn, major U.S. labs have accused Chinese companies of copying their models using a technique called distillation, where a large, powerful model tutors a smaller one.
“All of these Chinese AI labs were very cognizant from pretty early on that they were catching up,” said Kyle Chan, a fellow at the Brookings Institution. “Part of it is the appeal of being able to have this transparent, downloadable and customizable format.”
Because the internal parameters of open-weight models can be downloaded and modified, developers can customize them to a far greater extent than they can do for powerful closed models such as ChatGPT and Claude (The Post has a content partnership with OpenAI.). The software company Cursor, for instance, has used Moonshot AI’s Kimi K2.5 as the base for its most recent family of models specialized for coding. Starting from an open-weight model allows skipping the expensive step of teaching the machine basic language and going straight to more targeted training when making a specialized model.
But in Washington, the popularity of Chinese models has generated anxiety. In April, Republican lawmakers launched an investigation into Cursor’s then-parent company as well as Airbnb for their use of Chinese AI models, claiming the models posed a national security risk (Airbnb has used Qwen in its customer service agent, although it maintains it uses mostly American models).
And even among companies that use open-weight models, some still steer clear of Chinese ones. AT&T is among a growing number of large companies that use technology to automatically switch to a cheaper, less cutting-edge AI that is still useful for a particular task. Jeremy Legg, AT&T’s chief technology officer, said that the company will not use Chinese AI for national security reasons, but he added that the company’s rapid-fire AI switching technology does include other low-cost, open AI such as Meta’s Llama.
The open-weight era of Chinese models might not last forever, though. Alibaba has released several versions of Qwen that are closed and proprietary, and Reuters reported last week that Beijing was mulling restrictions on overseas access to homegrown models.
For many companies, open-weight models, regardless of where they’re made, are the most secure option.
“Being in charge of the weights means that nothing will ever change on you. No behavior will be different,” said Ahmad Osman, whose company builds custom AI infrastructure based on open models for companies. “Nothing will be seen by anybody outside of your institution.”
Shira Ovide and Kevin Schaul contributed to this report.
NewsletterDaily
Today’s Headlines
The most important news stories of the day, curated by Post editors, delivered every morning.