A growing number of US companies are testing artificial intelligence models developed in China as an alternative to systems from OpenAI and Anthropic. The main driver is cost, which has become a decisive factor for many businesses.
According to available data, Chinese open-weight models can cost up to nine times less than leading American models. This gap is reshaping the technology choices of several organizations.
Among the most cited providers are DeepSeek, Moonshot (Kimi), Zhipu (with its GLM model, also known as Z.ai) and Alibaba (Qwen). Most of these are open-weight models, which companies can run with greater flexibility.
A Price Gap of Up to Nine Times
The most cited comparison involves running the same workload across different platforms. Running that workload through Anthropic’s Claude costs about 4,811 dollars, while the same task on Zhipu’s GLM model costs roughly 544 dollars.
The difference amounts to nearly nine times the price. More broadly, open-weight models from DeepSeek, Alibaba (Qwen) and Z.ai (GLM) run 60% to 90% cheaper than flagship models from OpenAI and Anthropic.
Training costs are also part of the debate. China’s MiniMax reportedly trained its M1 model at a cost roughly 200 times lower than the figure stated for OpenAI’s GPT-4. (Source: https://www.aol.com/finance/china-minimax-debuts-m1-ai-165658590.html)
A Market Shift Toward Chinese Models
The change is also visible in usage data. Starting February 8, 2026, US companies routed more than 30% of their OpenRouter tokens to Chinese open models every week.
At some points the share reached 46%, compared with a 12-month average near 11%. The figure points to a rapid acceleration in the adoption of these tools.
A concrete example comes from Lindy, a San Francisco startup that builds AI assistants. The company moved part of its workloads from Anthropic to DeepSeek and reported savings of millions of dollars. (Source: https://cryptobriefing.com/openai-anthropic-pricing-pressure-chinese-ai/)
Performance and the Industry Response
Price is not the only factor at play. On the SWE-bench Pro benchmark for coding agent capabilities, Zhipu’s GLM 5.2 scored 62.1, beating OpenAI’s GPT-5.5, which reached 58.6.
Faced with this pressure, OpenAI is reportedly considering a significant reduction in token prices, according to reports from early June 2026. Such a move would suggest the company views Chinese price competition as a relevant factor rather than a marginal one. (Source: https://www.techrepublic.com/article/news-chinese-ai-models-cost-risk/)
Some questions remain for businesses, particularly those tied to data security and risk management when adopting models developed abroad. Some organizations choose to limit these models to non-sensitive tasks.
The overall picture shows a market in which price has become a central element of competition. For many companies, the potential savings are enough to justify at least a phase of experimentation.
Original article: cnbc.com
