American companies are quietly switching to cheaper Chinese AI models.
American companies are quietly moving real work onto cheaper Chinese AI models.
Fresh usage data reported by CNBC shows US firms shifting production workloads onto
open-weight Chinese models; Zhipu's GLM, DeepSeek, Kimi — as OpenAI and Anthropic
prices climb. On OpenRouter, the share of US-company tokens going to Chinese models
has held above 30% every week since February and peaked near 46%. On Vercel, GLM 5.2 was the fastest-adopted model of 2026:
daily token volume up roughly 27x and customer count up roughly 80x in its first
full week — landing within a point of Opus 4.8 at
about a fifth of the cost.
What it actually means
Yesterday's story was your frontier bill going up. Today's is where the money is quietly going instead — the counterweight to the whole 'AI billing is being rewritten this summer' thread. The 'good enough and 5-to-10x cheaper' tier stopped being a science project and became something US teams run in production; the token charts prove it, not the press releases. The move for someone building with AI isn't 'rip out Claude for GLM' — it's to stop treating your model as a fixed dependency. Route the boring, high-volume middle of your workload — classification, extraction, first drafts, bulk agent steps — to a cheap open model, and keep the expensive frontier model for the few calls where quality shows up in the output. The teams banking 60–90% aren't smarter; they just stopped paying premium rates for work a commodity model does fine. The cost that isn't on the sticker: data-governance and compliance questions around Chinese models — which is why this is a routing decision to make deliberately, not a default to flip.
Stop treating your model as a fixed dependency — route the boring middle to a commodity model and save the frontier price for the calls that show it.
46%peak share of US AI tokens going to Chinese models on OpenRouter
Source: CNBC · Wednesday 8 July 2026