Chinese AI Models Gain Ground in the US Market With Prices Up to 90 Percent Lower

Chinese language models are capturing an increasing share of the market among American companies using AI tools, leveraging a price advantage of up to 90 percent over offerings from Anthropic and OpenAI. Data from OpenRouter, the platform through which companies buy access to models from multiple providers at once, shows a sharp rise in the share of traffic generated by American businesses going to providers such as DeepSeek, Alibaba, and Z.ai.

Since February 8, 2026, the share of tokens processed weekly by American companies on Chinese models has stayed above 30 percent, reaching as high as 46 percent in peak weeks. That marks a jump from the 11 percent average of the previous twelve months, and just 4.5 percent recorded in the first half of 2025.

The main reason is price. DeepSeek V4 Flash costs $0.14 per million input tokens and $0.28 per million output tokens, while the pricier DeepSeek V4 Pro costs $0.435 and $0.87 respectively. That's a fraction of the rates charged by leading American models, where the cost difference for the same task reaches as much as ninefold by some estimates.

The Lindy Effect

The most telling example is Lindy, a startup building autonomous AI agents for businesses. Its CEO, Flo Crivello, announced moving all of the traffic previously handled by Anthropic's Claude models over to DeepSeek. The decision is expected to bring savings in the millions of dollars within the coming months.

Lindy is not an isolated case. Analysts also cite Canva and Vercel, the platform developers use to deploy applications, among companies increasingly turning to Chinese models, with traffic from new Chinese models growing faster there than for any other provider being tracked.

Price Versus Quality

The key development of recent weeks was the June 2026 launch of GLM 5.2, a model from China's Z.ai. In its first full week, the daily volume of tokens it processed grew roughly 27-fold, and the number of customers using it grew roughly 80-fold.

In one closely watched agentic-task benchmark, GLM 5.2 scored within less than a percentage point of Anthropic's Opus 4.8, while costing about one-fifth as much as the American model. The quality gap between top Chinese and American models is narrowing, while the price gap remains enormous.

Companies initially deployed AI regardless of which model was behind it, but now, as they scale up deployments, they're watching operating costs much more closely - Kyle Chan, Brookings Institution

Pressure on Anthropic and OpenAI

For Anthropic and OpenAI, the growing price competition presents a difficult choice. Both companies are investing billions of dollars in developing the most advanced models and the infrastructure to train them, making it harder to cut prices at the pace customers expect when comparing offers with Chinese alternatives. Justin Summerville of OpenRouter notes that the latest Chinese open-source models can already handle nearly all tasks except the most demanding workloads.

More companies are opting for a multi-vendor strategy instead of relying on a single model, matching the tool to the specific task. Where cost at scale is the primary concern, Chinese open-source models are becoming an increasingly common choice, while American models retain an edge where top precision or compliance with strict security requirements is essential.

What It Means for Polish Companies

For Polish companies using platforms like OpenRouter or working directly with providers' APIs, the price gap represents a real opportunity to cut AI deployment costs, especially for bulk tasks like document processing, customer service, or data analysis, where the quality of Chinese models' responses is often already sufficient. At the same time, choosing a model from China raises questions about data storage, compliance with RODO (Poland's implementation of the EU's GDPR data protection regulation), and regulatory risk, which is growing across the European Union as further AI Act provisions come into force.

Sources: IBTimes (ibtimes.com), CryptoBriefing (cryptobriefing.com)