With Chinese models nearly matching U.S. ones, prices a fraction of Anthropic's and OpenAI's have made cost a deciding factor for developers.
Openly released models from Chinese labs are now processing more than three times the weekly token volume of U.S. models on OpenRouter, the service that routes developers' API calls across a wide range of AI systems.
Price is what settles it. OpenAI and Anthropic charge a premium per token for access to their closed models, while DeepSeek, Moonshot, Zhipu and others publish weights that any developer can download and run with nothing owed per token. Chinese models come close to matching what their U.S. rivals can do at a small share of the cost. Zhipu lists GLM 5.2 at $1.40 per million input tokens and $4.40 per million output tokens; Anthropic's Opus 4.8 runs $5 and $25.
The cheaper option is not uniformly the weaker one. On the coding benchmark SWE-bench Pro, GLM 5.2 scored 62.1, ahead of GPT-5.5 at 58.6.
The turn is recent. Through most of 2025, U.S. systems took roughly 70% of usage among leading models. In the week of February 9 to 15, 2026, Chinese models moved 4.12 trillion tokens through OpenRouter against 2.94 trillion for U.S. models — the first weekly period in which the American side lost the lead on usage. In late June, all six of the most-used models on the platform were openly released systems from Chinese companies, among them Tencent, Xiaomi, DeepSeek, MiniMax and Z.ai.
The pattern shows up elsewhere in the plumbing. Vercel's Production Index, released in July, put open-weight models at 29% of the tokens passing through its AI gateway in June 2026, up from 11% in April. That is roughly one token in three for less than 4% of the money spent there. DeepSeek alone accounted for 22.6% of the gateway's token volume, third overall.
Kimi K3 from Moonshot AI follows the same logic — an open-source model the company says holds its own against leading U.S. systems, though across benchmarks generally it places behind both GPT-5.6 Sol and Claude Fable 5. On price it undercuts them: $3 per million input tokens, where GPT-5.6 Sol charges $5 and $30 and Fable 5 roughly $10 and $50.
Against the savings sits an exposure. Z.ai runs its API in the cloud, which places it under China's National Intelligence Law, so code or data passing through it could be reached by Chinese authorities. In May, U.S. legislators began examining what risks Chinese-origin models pose inside critical infrastructure.