Chinese artificial intelligence models from companies like DeepSeek and Zhipu are rapidly gaining ground in US enterprises, now accounting for more than 30% of weekly token processing on OpenRouter, a platform that lets developers access multiple AI models. This represents a dramatic shift from just 4.5% market share in early 2025, driven primarily by a widening price gap as US companies face mounting AI costs.

Why Are US Companies Switching to Chinese AI Models?

The core reason is straightforward economics. Open-source Chinese models cost between 60% and 90% less than leading alternatives from OpenAI and Anthropic, according to Justin Summerville, who works on data and analytics at OpenRouter. For enterprises processing millions of tokens daily, that difference translates directly to operating budgets.

Benchmarking firm Artificial Analysis quantified the gap by running identical evaluations across models. Running the same workload through Anthropic's Claude cost $4,811, while OpenAI's ChatGPT came in at $3,357. The same task through DeepSeek cost just $1,071, Moonshot's Kimi model $948, and Zhipu's GLM $544. This means Claude is nearly nine times more expensive than the cheapest Chinese alternative for equivalent work.

Enterprise AI spending is accelerating the urgency of cost control. Some 45% of companies surveyed by cloud cost firm CloudZero reported spending more than $100,000 per month on AI in 2025, up from 20% the year before. As those bills increase, procurement teams are scrutinizing where each dollar goes and whether every task truly requires a premium-tier model.

"Chinese AI models are particularly attractive to American companies now as AI costs skyrocket. Where previously U.S. companies were prioritizing AI adoption regardless of model, now they're getting more cost-conscious," said Kyle Chan, a fellow at the Brookings Institution's John L. Thornton China Center.

Kyle Chan, Fellow at the Brookings Institution's John L. Thornton China Center

How Are Enterprises Actually Using Chinese Models?

The adoption pattern is not a wholesale replacement of US models with Chinese ones. Instead, engineering teams are building intelligent routing layers that match each task to the cheapest model capable of handling it.

• Tiered Approach: A cheap open-source model handles routine work as the default, and when it encounters a task beyond its capability, it calls out to a premium frontier model from OpenAI or Anthropic for help, allowing companies to curb costs significantly.
• Task-Specific Routing: Engineering teams are beginning to route work to the cheapest model that is good enough for the job, rather than defaulting to the most capable option regardless of cost.
• Complete Migration: Some companies have gone further; AI startup Lindy moved 100% of its traffic from Anthropic's Claude models to DeepSeek, with CEO Flo Crivello telling CNBC the decision will save the company millions of dollars within months.

"Price is doing the work here. When a task doesn't need the best model, teams are beginning to route it to the cheapest one that's good enough, and the recent wave of models coming out of China is winning that trade," said Harpreet Arora, head of agentic infrastructure at Vercel.

Harpreet Arora, Head of Agentic Infrastructure at Vercel

This tiered approach is significant because it does not require Chinese models to match US frontier systems on every task. They just need to handle the lower-complexity work that makes up the majority of enterprise token volume. The premium models still get called, but far less often, and that alone is enough to reshape AI costs for large-scale deployments.

Are Chinese Models Actually Good Enough?

The cost advantage would matter less if Chinese models were significantly weaker. But recent releases have narrowed the performance gap to the point where the trade-off is increasingly favorable. Zhipu's GLM 5.2 landed within a percentage point of Anthropic's Opus 4.8 on one closely watched agentic benchmark, at roughly a fifth of the cost. Some researchers have said GLM 5.2 can perform on par with top US labs on certain cybersecurity benchmarks.

The adoption numbers reflect that performance improvement. GLM 5.2 saw the fastest adoption of any model tracked by Vercel in 2026, with daily token volume growing about 27 times and customer count growing about 80 times in its first full week after launch.

DeepSeek's V4 family, released in April, has had a similar impact. DeepSeek V4 Flash is priced at $0.14 per million input tokens and $0.28 per million output tokens, compared to $3 per million input and $15 per million output for Anthropic's Claude Sonnet 4.6 and $5 per million input and $30 per million output for OpenAI's GPT-5.5.

"The new open source models are performing well and prove capable for all but the most complex LLM tasks," said Justin Summerville, data and analytics lead at OpenRouter.

Justin Summerville, Data and Analytics Lead at OpenRouter

Brookings' Chan estimated that Chinese models are currently "six to nine months" behind top US rivals. That gap is real, but for many enterprise use cases, it may not matter. Most companies do not need frontier-level performance for routine tasks like data entry, customer support routing, or content summarization.

What Role Does Government Policy Play?

Most coverage of this trend frames it purely as a pricing story. But the shift toward Chinese open-source models is being accelerated by something that has nothing to do with price or performance: access and control. OpenAI announced recently that it would limit the rollout of its GPT 5.6 models because of a US government request. Anthropic had to pull its Fable and Mythos-class models after an order by the Trump administration, though export controls were eventually lifted after a tense standoff. During that 19-day disruption, developers who relied on those models had no access.

This created a supply-side push that reinforced the demand-side pull from lower prices. A model that can be downloaded, self-hosted, and run on a company's own infrastructure cannot be revoked by a government order. For engineering teams evaluating vendor risk, that permanence is now a factor in model selection.

"We're seeing companies increasingly motivated to turn to cheaper AI stacks they can control and adapt themselves, and given the state of open-source and open-weight models that often means leveraging Chinese options," said Yacine Jernite, head of machine learning at Hugging Face.

Yacine Jernite, Head of Machine Learning at Hugging Face

This creates an uncomfortable dynamic for US policy. Export controls designed to protect national security are pushing some US enterprises toward the very Chinese models that those controls aim to contain. The trend reflects a broader shift in how enterprises evaluate AI vendors, balancing cost, performance, and control in ways that were unthinkable just 18 months ago.