American firms are increasingly switching to Chinese AI models, and the reason is brutally simple: cost. As OpenAI and Anthropic hold premium prices, open-weight models from DeepSeek, Alibaba’s Qwen, Zhipu’s GLM and Moonshot’s Kimi have arrived at a fraction of the price while closing most of the quality gap. The result is a quiet migration that is now showing up in hard usage data — not just in opinion pieces.
Key takeaways
- US companies’ share of tokens spent on Chinese AI models via OpenRouter has stayed above 30% every week since February 8, 2026, peaking near 46% — up from roughly 11% a year earlier.
- The price gap is enormous: DeepSeek’s flagship runs around $0.87 per million output tokens versus roughly $25 for Anthropic and $30 for OpenAI.
- Named switchers include Lindy (100% to DeepSeek), Shopify (self-hosted Qwen 3), Coinbase (GLM 5.2 + Kimi 2.7) and Airbnb (Qwen).
- Reported savings range from about 50% to a 75x per-unit cost reduction.
- It is not only about price — open weights let companies self-host and keep data in-house, though they raise real governance and geopolitical questions.
How big is the shift to Chinese AI models?
The clearest signal is usage, not sentiment. According to OpenRouter data reported by CNBC, the share of tokens US companies route to Chinese models has sat above 30% every week since February 8, 2026, and has spiked as high as 46% — compared with an average of about 11% over the previous twelve months. In other words, close to half of some weeks’ US enterprise AI traffic now runs on models built in China.
The startup end of the market is moving fastest. Industry estimates suggest roughly 20-30% of startups now use open-source models, and about 80% of those pick a Chinese open-weight model. When a founder is watching runway, an order-of-magnitude difference in the AI bill is not a rounding error — it is the difference between shipping and shutting down.
The price gap driving the switch
The headline numbers explain the behaviour on their own. A flagship Chinese model can cost a small fraction of its US rivals per token:
| Model / provider | Approx. output price (per 1M tokens) |
|---|---|
| DeepSeek (flagship) | ~$0.87 |
| Anthropic Claude (flagship) | ~$25 |
| OpenAI (flagship) | ~$30 |
One widely cited workload comparison put the same job at roughly $4,811 on Anthropic’s Claude versus about $544 on Zhipu’s GLM — close to a 9x difference. Analysts broadly peg leading Chinese open models at 60% to 90% cheaper than the top US frontier models for comparable tasks. Before switching anything, it is worth modelling your own numbers rather than trusting a headline: our free AI API cost calculator estimates a real monthly bill from your token volume, and our AI price-performance index ranks models by intelligence-per-dollar so you can see exactly where each one lands on value.
Who is actually switching
This is no longer hypothetical. Several named companies have moved real production traffic:
- Lindy — the AI-agent startup moved 100% of its traffic off Anthropic’s Claude to DeepSeek, a switch its CEO expects to save millions of dollars.
- Shopify — reportedly replaced an OpenAI GPT-5 pipeline with a self-hosted multi-agent system powered by Alibaba’s Qwen 3, citing a roughly 75-fold reduction in per-unit language-model cost alongside higher output quality.
- Coinbase — cut AI spending by nearly half after moving workloads to GLM 5.2 and Kimi 2.7.
- Airbnb — runs 13 AI models but leans heavily on Qwen; CEO Brian Chesky publicly called it “very good,” “fast” and affordable. After rolling out a Qwen-backed customer-service agent, Airbnb says average resolution time fell from nearly three hours to about six seconds.
It is not just price: open weights change the calculus
Cost gets the headlines, but the second driver is architectural. Because these Chinese models are open-weight, a company can download and run them on its own hardware instead of calling someone else’s API. That flips two things at once: the per-token meter disappears, and sensitive data never has to leave the building. Airbnb, for instance, stressed that it does not send any data to the model developers. For teams weighing that trade-off, our self-hosting vs API calculator shows the break-even point where owning a GPU beats paying per token, and our open vs closed cost study quantifies how wide the gap has become. To compare specs, context windows and live pricing side by side, see our AI models database, and for a deep dive on the model leading this shift, our DeepSeek V4 guide.
The catch: governance and geopolitics
The switch is not free of friction. After Airbnb disclosed its use of Chinese open models, US lawmakers opened questions about the practice, even though the company self-hosts and shares no data upstream. For regulated industries, using a Chinese-origin model — even one running entirely on domestic servers — raises procurement, compliance and reputational questions that a spreadsheet alone will not settle. The pragmatic pattern emerging is to self-host the open weights (so no data crosses a border) and to keep a US frontier model on standby for the hardest tasks.
What it means for OpenAI and Anthropic
The pressure is already visible. OpenAI was reported in early June 2026 to be weighing sharp reductions in token prices — a move that would signal the company sees the Chinese price threat as existential rather than peripheral. The broader market mood has shifted from “tokenmaxxing” (throwing ever more tokens at a problem) toward efficiency: getting the same result for far less. That is precisely the environment in which a 60-90% cheaper model wins business, and it is why the next year of frontier-model pricing may look very different from the last.
Should your company switch? A quick framework
The honest answer is: sometimes. Run the decision on four axes. Volume — the higher and steadier your usage, the more a cheaper model (or self-hosting) pays off. Quality bar — for everyday drafting, extraction, classification and support, top open models are hard to distinguish from frontier APIs; for the very hardest reasoning, the US flagships still lead. Data sensitivity — if data cannot leave your control, self-hosting an open model is the cleanest answer. Governance — check procurement and compliance rules before committing. Model the money first with the calculators above, pilot on a non-critical workload, and only then move real traffic.
Frequently asked questions
Which Chinese AI models are US companies using most?
The most-cited names are DeepSeek, Alibaba’s Qwen, Zhipu’s GLM and Moonshot’s Kimi. DeepSeek dominates the price-driven switches, while Qwen has been adopted at Airbnb and Shopify and GLM/Kimi at Coinbase.
How much cheaper are Chinese AI models?
Analysts put leading Chinese open models at roughly 60% to 90% cheaper than top US frontier models. As a concrete example, DeepSeek’s flagship runs around $0.87 per million output tokens versus roughly $25 for Anthropic and $30 for OpenAI, and one workload comparison showed a near-9x gap ($544 on GLM vs $4,811 on Claude).
Is it safe to send company data to Chinese AI models?
Because these models are open-weight, companies can self-host them so that no data leaves their own servers — Airbnb, for example, says it sends no data to the model developers. The risk is less about data transmission and more about governance, procurement rules and geopolitics, which each organisation must weigh for itself.
Which US companies have switched to Chinese AI models?
Publicly named examples include Lindy (100% to DeepSeek), Shopify (self-hosted Qwen 3), Coinbase (GLM 5.2 and Kimi 2.7) and Airbnb (Qwen). Broader OpenRouter data shows US enterprise usage of Chinese models above 30% of tokens most weeks in 2026.
Do Chinese models match OpenAI and Anthropic on quality?
On many everyday and coding tasks they are now within a fraction of a point of the top closed models, which is why cost-driven switches make sense. For the very hardest reasoning problems, US frontier models still hold an edge — so a common pattern is to run a cheap open model by default and reserve a frontier API for the toughest jobs.
The bottom line
The migration to Chinese AI models is being driven by arithmetic, not ideology. When a capable model costs a tenth — sometimes a seventy-fifth — of the incumbent, and can be self-hosted to keep data in-house, cost-conscious teams will try it, and the OpenRouter numbers show many are staying. The lasting lesson is not “China won”; it is that AI inference has become a commodity where price and efficiency matter as much as raw capability. Companies that model their real numbers, pilot carefully, and match the model to the task will capture most of the savings without betting the business on any single provider.
Sources: CNBC, Forbes, Rest of World, Tech Startups. Reported July 2026.
