US Firms Adopt Chinese AI Models Amid Surging Costs

The Cost Crisis

Token prices for advanced models from US labs continue to climb, forcing companies to rethink their AI spending. Recent model releases from Chinese companies like DeepSeek and Z.ai are proving surprisingly competitive against leading frontier systems from developers like Anthropic and OpenAI.

Kyle Chan, a fellow at the John L. Thornton China Center at Brookings, noted that Chinese AI models are particularly appealing as AI costs skyrocket. The numbers back this up. The share of tokens used by US companies on Chinese AI models via OpenRouter, a platform for developers to access various models, has consistently stayed above 30% each week since February 8. This figure even reached 46% at times, a dramatic increase from the 11% average over the previous 12 months.

How Companies Are Making the Switch

The adoption process follows a clear pattern. Companies assess available options, integrate new models into existing systems, and optimize how they route different tasks across providers.

Step 1: Compare Capabilities and Pricing

The initial step for many firms is benchmarking the capabilities and pricing of available models. Chinese open-source and open-weight models are 60% to 90% cheaper than their leading American counterparts. While generally six to nine months behind top US rivals in overall capability, they perform well for most complex LLM tasks.

Take Z.ai’s GLM 5.2 as an example. It performed within a percentage point of Anthropic’s Opus 4.8 on one agentic benchmark, at roughly a fifth of the cost. For many workloads, that’s more than good enough.

Step 2: Integrate via Developer Platforms

Once a model is selected, companies integrate it into their existing infrastructure. Developer platforms like OpenRouter and Vercel facilitate this process, removing technical barriers to switching.

AI startup Lindy offers a revealing case study. The company moved 100% of its traffic from Anthropic’s Claude models to DeepSeek. Lindy CEO Flo Crivello stated this decision would save the company millions of dollars within months. DeepSeek, which had a bombshell release in early 2025, saw its share of gateway tokens climb sharply between May and June on Vercel.

Step 3: Route Tasks Strategically

Companies are becoming more strategic about how they route different AI tasks. Harpreet Arora, head of agentic infrastructure at Vercel, explained that teams are now directing tasks that don’t require the absolute best model to the cheapest option that is “good enough.”

This approach allows businesses to optimize costs without sacrificing necessary performance. Z.ai’s GLM 5.2, released in June, saw the fastest adoption of any model tracked by Vercel in 2026, with daily token volume growing approximately 27 times in its first full week.

The Real Numbers

What This Means

LaunchLemonade CEO Cien Solon noted that Chinese models like Z.ai and Alibaba’s Qwen are becoming viable options for specific workloads. These models offer an attractive combination of performance and cost that’s hard to ignore.

However, there’s a flip side. Yacine Jernite, head of machine learning at Hugging Face, highlighted a potential risk. Users might feel compelled to choose between expensive US proprietary models with fluctuating prices and limited accessibility, or Chinese models as the only feasible alternative for cost control.

The trend signals a broader re-evaluation of AI spending as companies mature their strategies. The days of defaulting to the most expensive frontier model may be ending. Instead, companies are building more sophisticated cost-conscious approaches that match model capabilities to actual workload requirements.