Something quietly dramatic is happening inside the software that US companies build on top of AI. Chinese AI models now account for somewhere between 30% and 46% of the enterprise API traffic flowing through US developer platforms, according to OpenRouter data reported by CNBC. A year and a half ago that figure was around 4.5%. The reason is not politics or novelty. It is price.
How far Chinese AI models have spread
The usage share for Chinese AI models has stayed above 30% every week since early February 2026, peaking at 46%, against a prior twelve-month average of just 11%. That is a fast, broad migration, and it is being driven by developers and finance teams rather than headlines. When the tools that quietly power apps, chatbots, and automations start running on different engines, the change reaches ordinary businesses whether or not they ever hear the model names.
The math behind the move
The cost gap is stark. As of mid-2026, DeepSeek’s V4 Flash was priced around $0.14 per million input tokens while a leading US frontier model sat near $5.00 for the same volume. Across the board, open Chinese models have been running roughly 60% to 90% cheaper than the top offerings from OpenAI and Anthropic.
At scale, that difference reshapes budgets. One widely cited example saw a major tech employer cap AI spending at $1,500 per engineer each month, then watch its full-year AI budget evaporate by April. When the same task can be done at a fifth of the price, the pressure to switch becomes hard to ignore. And the cheaper option is no longer obviously worse: one Chinese model, GLM-5.2, landed within a single percentage point of a top US model on a closely watched agentic benchmark. This is the same cost-versus-capability tension now driving the wider reshuffling of AI pricing tiers.
What it means for a small business
You do not need to pick sides in a geopolitical debate to take a lesson from this. The practical takeaway is that AI capability is commoditising fast, and paying premium prices for every task is no longer a given. For high-volume, lower-stakes work such as drafting, summarising, tagging, or first-pass research, a cheaper model, including many capable open-source options, may deliver most of the value at a fraction of the cost.
The caution is data governance. Analysts pairing the cost story with an “enterprise risk” warning are right that where your data goes matters, especially for anything involving customer records or sensitive business information. A sensible rule of thumb: route routine, non-sensitive volume to the cheapest model that clears your quality bar, and reserve premium or tightly controlled models for confidential or high-stakes tasks.
The bottom line
The story here is not really about one country’s models beating another’s. It is that the price of competent AI is falling faster than most budgets assumed, and the businesses that treat model choice as an active, reviewable decision, rather than a one-time default, will capture the savings. Match the tool to the task, keep an eye on where your data travels, and revisit the choice every quarter as the numbers keep moving.
