Chinese AI models are eating into OpenAI and Anthropic’s API business, but it's not a bad news for your business
Chinese AI models are gaining real usage among U.S. companies, and the reason is simple: price.
CNBC reports that Chinese models from companies such as DeepSeek and Z.ai have accounted for more than 30% of OpenRouter traffic every week since February 8. At times, that share reportedly reached 46%. Last year, the average was only 11%.
That is a big shift for a market still dominated in mindshare by OpenAI, Anthropic and Google. Developers may still prefer U.S. frontier models for the hardest tasks, but many AI workloads do not need the most expensive model available.
Chinese AI models are winning on cost
OpenRouter’s Justin Summerville told CNBC that Chinese open-source models are 60% to 90% cheaper. For startups running agents, customer support, coding workflows or automation products, that gap can decide which model gets used in production.
Lindy is one example. CEO Flo Crivello said the startup moved all of its traffic from Anthropic’s Claude to DeepSeek, saving millions of dollars while improving performance on some core workflows. The move also shows how quickly AI loyalty can change when inference bills become one of a company’s largest costs.
| Metric | Reported figure |
|---|---|
| Chinese model share on OpenRouter | Above 30% every week since February 8 |
| Peak share | Up to 46% |
| Last year’s average | 11% |
| Cost advantage | 60% to 90% cheaper, according to OpenRouter |
| Example customer shift | Lindy moved from Claude to DeepSeek |
DeepSeek is still behind the U.S. frontier, but the gap is narrowing
The U.S. still leads at the top end. The Center for AI Standards and Innovation evaluated DeepSeek V4 Pro in May and found that it lagged leading U.S. models by about eight months across areas such as cybersecurity, software engineering, natural sciences, abstract reasoning and math.
That gap may not matter for every customer. If a model is slightly behind the frontier but far cheaper, companies will route easier tasks to it and save premium models for the work that genuinely needs them.
That is why this is becoming a pricing problem for OpenAI and Anthropic. Their most powerful models can keep improving, but customers are learning to split workloads across multiple providers. Some requests go to the best model. Many others go to the cheapest good-enough model.
AI pricing is becoming the next big fight
The pressure is building because inference costs are no longer a side expense. AI agents can make repeated model calls, coding tools can burn through huge context windows, and enterprise apps can turn small prompts into large recurring bills.
That cost pressure is already visible in AI compute spending, where chips, data centers and power are becoming the hidden price behind every model launch.
Chinese AI labs now have a strong opening. They do not need to beat OpenAI or Anthropic on every benchmark. They only need to be good enough for enough workloads while staying much cheaper.
OpenAI and Anthropic can still defend the premium end of the market, especially for complex reasoning, coding, research and enterprise trust. The larger threat is margin pressure. If developers keep moving routine workloads to DeepSeek, Z.ai, Qwen, Kimi and other low-cost models, the API business becomes harder to price like a luxury product.
