American companies are increasingly turning to cheaper Chinese open-weight models like DeepSeek and Z.ai’s GLM 5.2 as token prices from OpenAI and Anthropic climb. Government limits on top US models have widened the opening. On some platforms, Chinese models now handle a striking share of US traffic.

Key Takeaways

• US firms are shifting to cheaper Chinese open-weight AI models
• Rising OpenAI and Anthropic token prices are driving the move
• GLM 5.2 nears Opus 4.8 on one benchmark at a fifth the cost
• Chinese models topped 46% of US token use on OpenRouter
• Government limits on US models widened the opening

Why US Companies Are Switching

The driver is money. As CNBC reported, advances in Chinese model capability arrive just as token prices rise at many US AI labs, leaving companies wrestling with unexpectedly high costs from using the technology.

The mindset has flipped. Where firms once chased the best model regardless of price, the priority now is efficiency, a shift the industry has started calling the move away from “tokenmaxxing.”

An analyst summed up the pull. Kyle Chan, a fellow at Brookings’ John L. Thornton China Center, told CNBC that Chinese models are especially attractive as AI costs skyrocket, noting that companies which used to prioritize adoption regardless of model are now far more cost-conscious.

The behavior follows the incentive. Engineers looking to build products and cut internal costs are increasingly testing cheaper open-source and open-weight models, and the most capable of those are made by Chinese companies.

The Models Leading the Charge

Two names dominate the story: DeepSeek and Z.ai. Both build open-weight systems, meaning parts of the model are exposed for developers to inspect, run, and sometimes modify, unlike the closed flagship models from OpenAI, Anthropic, and Google.

Z.ai’s GLM 5.2 is the standout. Released in June to major fanfare, it landed within a percentage point of Anthropic’s Opus 4.8 on one closely watched agentic benchmark, at roughly a fifth of the cost. It also posted the fastest adoption of any model tracked by Vercel in 2026.

Its strengths hit where enterprises care. GLM 5.2 is strong at agentic work like planning, coding, testing, and looping, the exact tasks companies are racing to automate, and some researchers say it performs on par with top US labs on certain cyber benchmarks.

DeepSeek remains a major force too. The company that stunned the industry in early 2025 launched a new model in April, and its share of gateway tokens on Vercel climbed between May and June.

The Lindy Case Study

One switch captures the trend vividly. In June, AI startup Lindy moved 100% of its traffic from Anthropic’s Claude models to DeepSeek.

The cost effect was dramatic. CEO Flo Crivello told CNBC the company watched the cost curve crash to the ground, and said the move will save Lindy millions of dollars within months.

Performance didn’t suffer either. Crivello said switching to DeepSeek V4 actually increased performance on many of Lindy’s core use cases, undercutting the assumption that cheaper means weaker.

The Adoption Numbers

The aggregate data backs up the anecdotes. On OpenRouter, a platform that lets developers tap a range of models, the share of tokens US companies run on Chinese models has sat above 30% every week since February 8, climbing as high as 46%.

The jump is steep. That share averaged just 11% over the previous 12 months and had fallen to 4.5% in the first half of 2025, making the current level a sharp break from recent history.

The leaderboard tells the same story. By some rankings, Chinese models now hold the top spots among the most widely used systems globally, with names like DeepSeek, MiniMax, Tencent, and Xiaomi collectively drawing more token traffic than major US frontier providers.

The Government Angle

Timing amplified everything. The surge in Chinese adoption lined up with Washington restricting access to the top American models, handing open alternatives a wide opening.

Both US leaders were affected. At the end of June, OpenAI said it would limit the rollout of a new set of models at the government’s request, while export controls on Anthropic’s Mythos and Fable were lifted that same month after a tense standoff with the Trump administration. Anthropic has said that suspension came from US export controls, which the Commerce Department later cleared.

That backdrop reframed what buyers value. A model that can be downloaded and run on a company’s own servers can’t be revoked by anyone, and amid federal oversight, that permanence started to look like the safer institutional bet.

What It Means Going Forward

The momentum is real, but so are the caveats. Even as Chinese models close the gap, a capability lead for US labs hasn’t vanished.

DeepSeek itself offers a sober read. The company estimates Chinese models trail leading US systems by roughly 3 to 6 months in pure capability, a reminder that price, not raw performance, is doing much of the persuading.

There’s a deeper worry for US labs. Anthropic has alleged that operators linked to Alibaba’s Qwen lab engaged in distillation to extract knowledge from its models, and the fear is that if distillation spreads, Chinese firms could close the gap quickly even as US companies spend heavily on frontier research.

The policy fight is just beginning. As Washington weighs how to regulate its most powerful models and how to slow the adoption of foreign alternatives, US companies are quietly voting with their budgets, and right now a growing share of those dollars is flowing toward cheaper stacks they can control themselves.

Digital Trendings is your trusted source for AI news and updates, stay tuned for more.