What happened
The share of AI work that U.S. companies run on Chinese models has quietly become the biggest cost story in the industry. Through OpenRouter, the platform developers use to send tasks to different models, U.S. traffic to Chinese systems has held above 30% every week since February and spiked as high as 46% — up from an 11% average over the prior year.
The pull is price. Open Chinese models from DeepSeek, Minimax, Moonshot, and Z.ai run 60–90% cheaper than the leading Anthropic and OpenAI systems, and in June, Z.ai's GLM 5.2 saw the fastest adoption of any model tracked by Vercel this year.
American businesses aren't switching to Chinese AI out of loyalty — they're doing it because most of the work they hand a model doesn't need the most expensive one.
The detail almost everyone will miss
The headline reads like a national-security story. For an operator, it's a pricing story — and the number that matters isn't the country, it's the gap.
One developer told Rest of World an hourlong coding session cost about $10 on Claude and under 50 cents on the Chinese model doing the same work. Another pays $500 a month for Claude and ChatGPT to handle his hardest planning tasks — and $200 a month for cheaper models that cover the other 90% of the workload.
The winning move isn't picking one model; it's routing each task to the cheapest one that's still good enough — and the "good enough" tier now covers the overwhelming majority of real work.
Why this matters if you run a business
If you pay for AI by usage — and almost every serious tool now does — your bill is not a fixed cost. It's the result of a decision you probably haven't made on purpose: sending routine work to a premium model because it was the default.
This is not a call to run your company on Chinese AI. For most regulated or data-sensitive businesses that's the wrong trade, and it's already drawing Congressional scrutiny of the firms that did it. The transferable insight is the tiering itself.
The same task quality can cost you ten times more or ten times less depending only on which model you route it to — and the frontier providers know it, with OpenAI now weighing steep price cuts to keep enterprise customers from leaving.
What to do about it
You don't need to change vendors to capture most of this. You need to stop treating every task as equally hard:
- Tier your AI work. Sort the jobs you hand to AI into two piles — the genuinely hard reasoning (contracts, strategy, edge cases) and the routine bulk (drafting, summarizing, formatting, first-pass code). The second pile is usually 80–90% of the volume.
- Route the bulk to a cheaper model. Point the routine pile at a lower-cost model that clears your quality bar, and reserve the premium model for the hard 10%. Most platforms let you set this per task without touching your data policy.
- Draw the security line first. Decide what data may touch which model before you optimize for price — run sensitive work through a U.S. provider or your own environment, and let cost decide only among options you already trust.
The companies getting AI economics right in 2026 aren't the ones on the best model — they're the ones who stopped paying frontier prices for grunt work. The savings were always there. Someone just has to make the routing a decision instead of an accident.