As is well known, US AI companies - OpenAI, Anthropic, Google, do not provide services to China. But in such a blocked region, the price people pay to access Claude/GPT is lower than elsewhere, even less than one-fifth of the official API pricing. How is this possible?
I want to share a complete perspective here: how an ordinary Chinese student, also developer, accesses the most advanced US models. Including the evolution of various access methods over 4 years, the practical experience of using various methods, and the problems encountered, etc. I will try to describe it objectively and truthfully.
Let's start by going back to November 2022, when OpenAI released "ChatGPT":
Phase 1: "ChatGPT"
ChatGPT was released, and it was big news in China, even though it predictably did not serve China. Even though I was still a high school student at the time, I was still interested, after all, it was the first time I saw something truly close to "intelligence".
How to access it?
There are two types of services generally inaccessible in China: one is that the GFW blocks the domain name or IP of the service, and the other is that the service provider refuses IPs from China. ChatGPT is both. The solution is also simple, use a proxy, which is a basic skill for Chinese developers.
In addition, registration requires receiving a SMS verification code. Chinese mobile numbers are definitely not an option, but the solution is not difficult either, find a verification code receiving platform, use a temporary number to receive the verification code.
Thus, I started using ChatGPT, which now seems like a model that speaks slowly and is not very smart.
Phase 2: "Mirror Sites"
During high school, I didn't have many scenarios to use ChatGPT. I started using AI more when I entered university, as AI is well-suited for dealing with those annoying assignments.
It was the second half of 2023, and there was a new way to access it: mirror sites.
"Mirror sites" originally referred to another type of website - software source mirrors. Due to China's network environment, almost all software sources of package managers are unavailable. Therefore, some organizations have set up mirrors of these software sources in China, and you can download them at high speed by simply modifying the package manager's software sources.
The idea of ChatGPT mirrors is similar, its interface is almost a copy of the official ChatGPT website design; but at the bottom, it calls the OpenAI API through a proxy server to generate chat completion.
The advantages of this method are obvious: users no longer need to configure proxies or have foreign mobile numbers. But since it uses the API, an API Key is required, and users are usually prompted to enter an API Key upon first access.
So where does the API Key come from?
At that time, OpenAI gave new registered users a $5 credit, so some people registered accounts in bulk, obtained API Keys, and then sold them. The places where they were sold were usually "card issuance websites" set up by themselves. Such websites specializing in selling virtual goods have existed before, and with existing templates, setting them up was not difficult.
The industrialization of this industry began at this time:
- Upstream bulk creation of accounts, obtaining API Keys and selling them
- Due to information opacity, these Keys may be resold and marked up multiple times before being sold to users
- Some card issuance websites even have built-in functions for opening sub-sites and earning commissions
- Although mirror sites themselves do not make money, they can earn advertising fees by linking to card issuance websites, or sell API Keys themselves.
The selling price of these API Keys is far below $5 – they have almost no cost, and the selling prices vary greatly between different merchants. The seller I found initially sold them for ~1.5 yuan each, but I later found merchants selling them for only ~0.5 yuan each.
However, this method still has many problems:
- These API Keys can only use gpt-3.5, without gpt-4 access.
- These API Keys have a rate limit of 2 RPM.
- The $5 credit can be used for a long time with gpt-3.5, but sometimes the account is banned before you finish using it, and the API Key becomes invalid.
- Merchants often provide one or two days of warranty, but these verbal promises are not guaranteed.
The biggest problem should be the inability to use gpt-4, as the improvement of gpt-4 over gpt-3.5 is too great! Fortunately, someone has researched other methods:
At that time, Microsoft provided gpt-4 in Copilot Pro. So someone developed a service to reverse engineer Copilot's API and provide its model in OpenAI API format. I happened to have Copilot Pro (from GitHub Student Pack), so I deployed this service. Then, by modifying the base URL and API Key in the mirror site to the corresponding values, I could use gpt-4.
This situation did not last long. Around the first half of 2024, one day, I found that all the API Keys I bought from a merchant had become invalid. When I entered the merchant's website again, I found that they were no longer selling API Keys, but linked to a completely different website. At this time, I realized that there was a completely new model.
Phase 3: "Relay Stations"
This new website does not sell API Keys, but is a complete API platform, similar to OpenRouter (although I didn't know about OpenRouter at the time). You can create an account, top up, create an API Key, and use it. It offers both gpt-3.5 and gpt-4 series models, just like the official API, billed by usage, but at a lower price. Later, I learned that such websites are called "relay stations", where Tokens are relayed before being sent to users.
These types of websites initially only offered gpt models, but with the development of models like Gemini and Claude, they also began to offer more models. And they also did what OpenRouter did – format conversion, allowing users to call various models in a unified OpenAI-compatible format.
There are several sources for these relay station APIs. The simplest is to use a US IP and a US credit card to top up normally on the official API platform, and then relay its API to the relay station. This channel is the most stable, but the price is more expensive than the official one, usually 8 yuan per dollar.
Secondly, there are "official transfer" channels. These channels often obtain bonus credits by bulk registration and binding cards on the official platform, and use these bonus credits to call APIs. The prices of these channels are often about 1/3 of the official prices. Among them, Claude channels are slightly more expensive than GPT channels; Azure channels are lower than OpenAI, AWS is lower than Anthropic, and Gemini channels are close to Azure. This type of channel was the most used at the time, and in terms of functionality, it was basically no different from the official channel, but its availability was slightly lower.
Finally, there are reverse engineering channels. They reverse engineer ChatGPT web pages and other chat UI and provide them in API format. This type of channel has the lowest price, but is often billed by the number of calls rather than Tokens. However, this type of channel has many limitations. Since it reverse engineers the chat UI, it cannot modify system prompts and model parameters, nor does it support features like tool calling. In addition, the APIs they construct are often not standard. If you use ai-sdk to call such interfaces, Zod errors may be thrown in some cases because the response format cannot pass strict validation.
A relay station often provides multiple groups, each group corresponding to a channel. When creating an API Key, you can specify a group to call models through a specific channel.
At this stage, the clients we used also changed. Mirror sites were gradually eliminated and replaced by open-source universal AI chat clients. Typical representatives include NextChat, LobeChat, OpenWebUI, etc. The design intention of these clients is to call models through official APIs, but since relay stations provide OpenAI-compatible APIs, you only need to change the baseURL to use these clients through relay stations.
BTW, I also developed an open-source AI chat client during that period, but its star count is not very high, you probably haven't heard of it.
At this stage, it seems quite perfect, right? You can access all models in one place, choose different priced channels according to your needs and budget, billed by usage, and provided in a unified OpenAI-compatible API format, compatible with all mainstream clients. This seems to be the ultimate solution.
In fact, it's mostly true. From the beginning of 2024 to the present, relay stations have been the main way for Chinese people to access US models. Only the release of Claude Code has brought about significant changes in this industry.
Phase 4: "Coding Agents"
Claude Code was quite popular in China when it was released. But after I heard about it, I didn't pay much attention. I thought: Isn't this just a slightly better Cursor / Copilot, and even worse than Cursor / Copilot in terms of IDE integration?
So I didn't try it, let alone consider its impact on the relay station industry. But as it turned out, new application formats like this have a huge impact on the industry. Usability is secondary; the main thing is its billing. Previous applications, such as Cursor, had relatively normal usage allowances in their subscription plans. So the way they were used in China was mainly by repeatedly registering for free trials or by connecting to relay station models to reduce costs.
But Claude Code is different. Its subscription plan usage limits are far higher than API of the same price – only model vendors can achieve this. And calculated by usage limits, its price is even lower than that of relay stations, which changes the game rules.
At this time, there was no convenient way to use Claude Code in China – purchasing a subscription yourself requires bypassing layers of restrictions, and relay stations mainly provide OpenAI-compatible API, not compatible with Claude Code; Agentic coding greatly increases token demand, and its subscription plan usage is so generous. This is clearly another huge business opportunity, another one I missed.
And the products that emerged from this demand are relay stations specifically designed for Claude Code.
The working principle of such relay stations is roughly that the upstream connects to a pool of Claude Max subscription accounts, reverse engineers the Claude Code API, and connects it to the relay station as the upstream. The relay station also provides the same API to downstream users, who only need to modify the base URL and auth token in the Claude Code configuration file to use the relay station.
The difference between these relay stations for Claude Code and previous relay stations is:
- Lower Price: The price of this reverse-subscribed channel is lower than that of the official transfer API channel (by this time, Claude official transfer channels had already increased their prices).
- Cache Optimization: Prompt caching is particularly important in vibe coding scenarios. These relay stations often implement channel stickiness to hit the cache.
- Dedicated rather than General Purpose: This channel often only allows Claude Code to be used and does not allow external connections for other purposes, to avoid triggering risk controls.
- Billing Method: Some are pay-as-you-go based on usage, but some are subscription-based to match Claude's subscription plans.
Since Vibecoding consumes significantly more Tokens than Chatbots, and is highly correlated with productivity, users' willingness to pay is high; such relay stations have developed rapidly. Those who were quick have already made a fortune.
After this, traditional relay stations also wanted to enter this market, so they started adding support for Claude APIs, adding reverse subscription channels, and began optimizing cache hits.
With the release of more Coding Agents like Codex and Gemini CLI, relay stations for Claude Code began to support other models and API formats. Some channels also began to allow external API connections to expand their uses. Even due to the difficulty of managing subscription plans, they have completely switched to pay-as-you-go billing based on usage.
Thus, the two types of relay stations have undergone "convergent evolution" and become the same thing.
Problems
These relay stations allow Chinese people to access the most advanced US model APIs without proxies, and at discounted prices. So, what is the cost?
Availability
The biggest problem is availability. Generally, channels that are more expensive than the original price (official channels) have few availability issues. But we use discount channels the most. As mentioned above, these channels usually come from bulk gift credits and reverse engineering. Once these accounts trigger risk controls and are detected, they will be banned. A relay station has a pool of accounts, and if a large number of accounts are banned in a short period and cannot be replenished in time, there will be a shortage. In practice, this means that requests sometimes fail with errors, and sometimes TTFT is very long. There are even times when it is impossible to use smoothly for several consecutive hours. Gemini models may also result in empty replies, which are considered successful responses according to the protocol and are even billed. This situation cannot be solved by simply configuring fallbacks and timeouts.
Privacy
When using such relay stations, all your requests and responses will inevitably be processed by the relay station. These relay stations themselves are non-compliant, usually operated by individuals or small teams. Some do not even have privacy policies, and even if they do, you cannot expect them to necessarily comply. There are rumors that some relay stations sell dialogue data to domestic LLM companies for training data. Although there is no definitive evidence, no one can guarantee that this will not happen. As for more serious actions, such as stealing sensitive information like API Keys from dialogue data, I can only say that it is possible, but less common. Because this method is more complex, more likely to offend users, and difficult to obtain stable output; I have not heard of explicit cases.
However, if you are willing to pay a higher price than the original, there are better options. OpenRouter was considered the "largest relay station" by us for a period of time. It was much more reliable than other relay stations. Although a proxy is still required, you can register and top up to reliably use any model. But good times don't last long. In the second half of 2025, perhaps model vendors also discovered that a large number of Chinese people were accessing their models through OpenRouter, so OpenRouter also added restrictions. The restrictions mainly targeted OpenAI, Anthropic, and Google models. First, requests from Chinese IPs, including Hong Kong, were unable to access these models. Later, as long as an account used a Chinese payment channel (including UnionPay cards, WeChat, Alipay) once, that account would never be able to access these models. Although there are still some OpenRouter alternatives without these restrictions, not many people seem to be using them.
Deception
This problem is more common in the sinking market, such as merchants on Taobao and Xianyu. Their methods include replacing models and secretly altering the multiplier. I once bought their tokens out of curiosity on Xianyu – they claimed the price was much lower than that of relay stations. But soon, I found the problem: the token usage recorded on their website was several times higher than my actual usage. I raised questions, they did not admit it, but were willing to give me a refund. On e-commerce platforms, this is important to reduce disputes. They make money from users who don't even discover this.
But truly professional users would never buy here. Reliable relay stations are independent sites, and they generally do not engage in model swapping or secretly altering the multiplier. The reason is simple: their actual costs are far lower than the prices they sell to users, and the cost of acquiring new users is higher by comparison. Their non-compliance also reduces promotion channels. They need to retain users rather than reduce costs. E-commerce platforms, on the other hand, have continuous traffic, but prices are transparent, and users don't know which is real or fake, so they choose the cheapest one. Therefore, they will try their best to reduce prices.
Examples
I will use two real-world examples — two relay stations I’m currently using — to illustrate the two different types of relay stations.
The first is an upstream relay station, which explicitly sells API from different sources. The price of this type of transit station is usually lower, but the availability is lower and it is more complex — it requires understanding the specifics and limitations of each channel.
The second is an aggregation relay station, which aggregates various channels internally, thus has higher availability. It no longer distinguishes channels to the outside and usually provides universal API, but the price will be slightly higher.
From a Broader Perspective
Let's return to the starting point, the root of the problem – the two-way decoupling between China and the US in the AI field, and consider the logic behind it.
US AI companies blocking services to China may be due to:
- Compliance. China has strict regulations on generative AI, and US export restrictions on China's AI industry are becoming increasingly strict.
- These companies may also want to use the blockade to prevent China from learning such technologies to maintain their monopoly.
China also blocks these services through the GFW, which may be due to:
- Security. The Chinese government is concerned about a large amount of data being processed in the US, and also worried that AI-generated content will be controlled by the US.
- Blocking US AI services can promote the development of local AI industries.
There are obviously contradictions here. Does such a blockade hinder or promote the development of China's AI?
As mentioned above, such a blockade has not completely prevented China from accessing advanced AI. Professional users always have ways to access it, which does not prevent relevant industry personnel from observing and learning from these models and products.
For ordinary users, they indeed cannot access these services, but this leaves a huge market. As long as you can be the first to create a compliant, local AI and put it on the app store that ordinary users can easily access, you will inevitably gain a large number of users and seize this huge market.
From the result, Chinese LLMs have emerged one after another, gradually approaching the cutting edge. Except for closed-source US companies, all flourishing LLMs come from a blocked region, while other unblocked regions have no models to speak of. The answer to this question is self-evident.
Looking Ahead
So, what will the future hold? Will these relay stations continue to exist? My view is yes.
US AI companies naturally know that people are stealing their computing power, and they are constantly improving verification and risk control mechanisms; but relay station operators are also accumulating experience and becoming more proficient in this process. This is like an endless arms race.
In fact, this cycle has been happening for the past two years:
- Service providers add new verification/risk control mechanisms, a large number of accounts are banned, relay station prices increase, and availability decreases.
- Relay stations become familiar with verification/risk control mechanisms, or find new channels, prices decrease, and availability increases.
Now, it may be at a time when prices have just passed the low point and are gradually rising.
Absolute blockade is impossible. There is no wall across the Pacific. This dynamic balance will continue.
AI in China and the US will continue to develop. Recently, the progress of open-source models has slowed down, and the slight lead of US AI may continue for some time. China's AI may have already surpassed in some areas, such as video generation. I still don't understand why OpenAI gave up Sora, which almost handed the entire field to ByteDance.
In any case, this is very interesting. Let's see what happens next.
