Chinese Open Models Cut Inference Costs 60-90%: The Catch Is Where the Data Goes

What Happened

As OpenAI and Anthropic API prices climb, US and European companies are quietly moving to Chinese open-weight models. On the routing service OpenRouter, the share of tokens US companies send to Chinese models has topped 30% every week since February and hit as high as 46%, CNBC reported, against an average of roughly 11% over the prior twelve months. The reason is blunt: open models like DeepSeek, Alibaba’s Qwen, and Moonshot’s Kimi handle the same work 60% to 90% cheaper than the frontier models from Anthropic and OpenAI.

The gap gets sharper up close. As Rest of World detailed, an hour of coding runs about $10 on Claude versus under $0.50 on DeepSeek. Lindy, a San Francisco AI-assistant startup, switched from Anthropic to DeepSeek in June and expects to save millions a year, with no meaningful quality drop on core work like email triage and automated replies. On the deployment platform Vercel, DeepSeek’s token share jumped from under 1% to 17% in a single month in May, and Airbnb and Cursor (Anysphere) have both disclosed using Chinese open models like Qwen and Kimi. The price war is escalating too. DeepSeek cut V4 Pro API pricing by 75% to $0.435 per million input tokens and $0.87 per million output tokens, a quarter of the earlier $1.74 and $3.48, InfoWorld noted.

What This Means for Founders

Be honest first: this is real cost savings. For latency-tolerant, high-volume work, think document classification, summarization, first-pass replies, and overnight batch jobs, there is less and less reason to insist on a frontier model. But if you chase the price alone, there is a bill leaking out the back.

The core tradeoff is where your data flows. Point at an API hosted in China and your prompts, documents, embeddings, logs, and telemetry cross your company perimeter into another jurisdiction, as InfoWorld flagged. For an enterprise-selling founder that is not a vibe, it is procurement. Your SOC 2 and GDPR posture, your data processing agreements, and a regulated customer’s security review all hinge on where processing happens and what is retained. So two workarounds have settled in. One is pulling the open weights and running them on your own servers or a sovereign or in-region cloud. The other is going through a US cloud provider so you never pay a Chinese developer directly and the data never leaves your region. Regulated-industry firms are already choosing the US-cloud route over paying Chinese developers directly, Rest of World reported.

Political risk belongs in the cost line too. In the US, congressional inquiries have opened into Airbnb’s and Cursor’s use of Chinese models. If geopolitics shifts, a model that was cheap yesterday can become awkward to run tomorrow. So rather than hard-wiring your code to one model, put an abstraction layer between your app and the provider so you can swap models per task. Because these are open weights, the same model runs identically across many providers, and that is your hedge.

What You Can Do Now

Split your workload into two buckets. Route the batch work that carries no sensitive data and tolerates latency to a low-cost Chinese open model as a candidate, and measure real quality and per-unit cost side by side against a frontier model. Open the sensitive-data work only through self-hosting or an in-region path, and settle your cross-border transfer consent and log-retention policy before you turn it on. Either way, keep your payment path and your data path logged separately, so when a customer’s due diligence or a regulator’s review lands, the answer is already written down.