This year’s Singles Day shopping carts in China may include a new category: tokens. Alibaba Cloud, DeepSeek, Zhipu, MiniMax and Kimi have all moved onto e-commerce platforms in recent weeks, and AI subscriptions and token top-ups have become a promotional genre of their own. Tmall is reported to be designing bundles aimed at students, families and one-person studios.
Behind that consumer surge, the layer that sits between model providers and users is dividing into winners and casualties. Demand is growing. Not every business that sells tokens is still making money.
A 71 per cent margin on routing
OpenRouter is the clearest example of the aggregation model. The platform carries a 71 per cent gross margin. In August, payments group Stripe agreed to acquire it. Neither side disclosed the price, but press reports put the deal above USD 8 billion. Stripe said OpenRouter connects more than 400 models across more than 80 suppliers.
What the buyer paid for is not token inventory. It is the decision layer that lets a developer pick between models on capability, price, speed and reliability for each task. A company can manage calls through one platform and route dynamically instead of integrating every model separately.
China’s nearest equivalent is SiliconFlow, which sells serverless token services, dedicated instances, inference acceleration and private deployment. It connects models, compute and system software so that a model can run on different vendors’ GPUs and AI chips, then delivers inference as token services or as technical deployments.
As of June, SiliconFlow reported more than 10 million registered users and more than 13,000 enterprise clients, with average daily throughput of about 578.5 billion tokens and a single-day peak above 1.07 trillion. It has completed seven funding rounds, reached a valuation of RMB 7.74 billion and filed to list on the Hong Kong Stock Exchange.
The reseller layer underneath
The API reseller is a different animal, and it has no clean definition. Small operators aggregate model endpoints, payment convenience and price gaps to pull in users, offering prepaid balances, a unified interface or bundled quota, usually aimed at getting around access restrictions on overseas models.
The discounts are not subtle. On one reseller’s price page, Claude Fable 5.1 is quoted at USD 1.5 per million input tokens and USD 7.5 per million output. The official standard price is USD 10 and USD 50. That is roughly 15 per cent of list, and comparison sites have sprung up to rank platforms by quote, exchange rate and top-up cost.
A second model goes further. Some operators sell a finished account bundled with a subscription, for example RMB 150 a month for an aged Claude account plus a Pro plan, with a guarantee on first login and subscription activation but no guarantee against a later ban. The operators are candid about the reason. Too many accounts are being closed in China.
Anthropic’s own transparency data shows the scale. Between January and June 2026 the company banned 11.4 million accounts and received about 398,000 appeals, of which roughly 42,000 were overturned. A market that supplies access at 15 per cent of list price cannot exist without that enforcement gap.
What is actually behind the discount
Resellers attribute cheap supply to a mix of upstream channels: volume purchasing, vendor promotion credits, or buying subscription plans and splitting the quota inside them for resale. Those are the respectable explanations.
The less respectable one is the model itself. On a reseller platform, the model name in the listing is filled in by the operator. The endpoint behind it may not be the full-weight model at all. Users have posted repeatedly about performance falling off a cliff when a task gets hard.
This is not only user suspicion. In March, a research team at the CISPA Helmholtz Centre for Information Security published an audit of shadow APIs titled Real Money, Fake Models. It found output that diverged materially from official endpoints, with performance gaps up to 47.21 per cent, and model fingerprinting failed in 45.83 per cent of tests.
Inside reseller communities, model substitution is an open secret. Detection tools and purpose-built prompts now exist to test whether a service is what it claims, and cautious buyers run a test before topping up.
For anyone building on top of these services, the risk is not the price. It is that a production system can be silently downgraded, and the invoice will never tell you.
Editor’s note: This is an adapted translation of the original Sohu Tech report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.sohu.com/a/1093248985_122014422.