In Anthropic's latest barrage of AI doomerism and China hawkism, they claim that Moonshot and DeepSeek are routing their users' requests to Anthropic models instead of their own models. I guess what Anthropic is claiming as the reason is that their models are not strong enough so they have to route to Opus to produce good results.
This is false, and Anthropic knows that it is false. It is also illogical, and a much simpler explanation exists.
Users in China cannot sign up for Codex or Claude accounts; the sign up flow simply does not allow for Chinese phone numbers to be used for verification. Instead, Chinese users figure out a way to get a foreign phone number and register that way. However, these users will frequently get banned simply for being detected as Chinese users.
An alternative method for accessing these models exists, the "transit station" as it is called in Chinese. In English, we would call it an AI gateway.
This is a subscription to Opus and Fable sold by a transit station on Xianyu, China's equivalent to Ebay.
Transit stations exist to solve the account creation issue, and also to make it possible to use the models without a VPN.
China is a very price sensitive market, so transit stations need to offer extremely cheap prices. They may advertise a ChatGPT subscription for just $1 or in that range.
There are a few different ways of reaching these pricepoints. Several months ago, the transit stations would exploit bugs in the Android payment process where they got the ChatGPT subscriptions extremely cheaply. There has also been certain promotions or education discounts with extremely low prices that have been exploited, and there's also been geo-arbitrage, where accounts are registered in say India or Mexico, due to their significantly lower subscription fees.
These subscriptions are then pooled together and exposed as an API that the customers of the transit stations access the models via, routing between subscriptions as they hit their 5h and weekly limits.
But these are just one of several ways for the stations to profit off of "ChatGPT subscriptions" and "Claude subscriptions" that are too cheap to be true.
Another way to make money as a station is to store all of the traces that are routed and sell them to labs as training data. I would guess that every lab in China buys these traces, and they may even be sold outside of China. This is a massively larger revenue stream than any subscription fee, and is how they can sell model access for almost free.
This researcher bought such data, and was able to extract credentials from users who were using transit stations to access language models. He calls the transit stations "LLM routers" but they are not regular routers, and OpenRouter is not one of them. These are all Chinese transit stations:
A third way to profit, or perhaps rather cut cost, is to route to cheaper models. How are users going to know if they're really talking to Opus? So, while transit stations claim to give you access to Opus, they are actually routing you to cheaper models like Kimi, DeepSeek and GLM.
And there you have it. The reason Anthropic is seeing Kimi and DeepSeek traces is because these gray market model gateways are scamming their own users.
Anthropic certainly knows this, but as always, they choose to lie, cheat, and deceive you.