After Claude's KYC: Relay Stations Aren't Dead, But It's Harder for Ordinary Users

Original Author: Kaori

Mo is the operator of a Claude relay station. In less than two months, she has served a group of clients both domestically and internationally who have high demands for Claude model performance. Among them, 80% are in China, with the rest scattered across Europe, Latin America, South Asia, and the Third World—"notably, none from the United States."

She has also shared her observations about this business on her webpage and social media accounts, introducing many users who were previously unaware of relay stations to the workings of this gray area. A relay station is an intermediary role that constantly calibrates between platform risk control, technical capability, and client trust.

Mo describes herself as not a typical relay station operator. She doesn't engage in price wars and only provides high-quality services to B2B clients. She has a bold theory: the prerequisite for this business's long-term existence is the continued dominance of American models. However, she is simultaneously pessimistic about AI overall. Even with regional pricing discounts of 40%, high-quality intelligent resources remain a heavy burden for many people in Third World countries.

Yesterday, news broke that Claude was introducing passport KYC verification. Initially, Mo judged that this wouldn't have a huge impact on the relay station industry. But just one day later, the chill of KYC has already been felt. What concerns her more is that this new rule might push more ordinary users towards relay stations, while the cost for ordinary users to identify reliable channels is too high, which might not be a good thing.

The following is Mo's account:

When the news about Claude requiring passport KYC verification came out yesterday, many people thought relay stations would be completely finished. But currently, Claude's KYC is randomly triggered. I initially thought my business wouldn't be significantly affected for the time being.

The next day, I still found that this incident had a negative impact on our operations. The overall cost of the account pool will increase, and channel stability will also be affected to some extent.

Whether it's stockpiling accounts or finding KYC suppliers, the entire relay station business has many countermeasures. Of course, this is just for now. It's hard to say if it will later be extended to all users.

Claude implementing KYC is its own choice. The entire AI field is still in an upward phase. Other companies haven't faced the same level of demand or other pressures as Claude, so other models haven't implemented KYC yet. But the future situation is uncertain.

The Business of Relay Stations

The relay station business is essentially buying low and selling high—extracting account quotas, selling them out within a certain payback period, and then making a profit.

The barrier to entry in this line of work is really low. As long as you have one or two clients, you can start.

I once encountered someone who sold at quite high prices. I went to check his website and found that after six months of operation, the URL was still HTTP. The defense rate against attacks is very low, and client data is easily leaked.

Furthermore, operating this business cleanly is not everyone's choice.

One type involves packaging users' call logs. Some relay stations, to boost their C-end volume with seemingly unbelievably low prices, will package and sell backend database information—like the prompts you input and the content returned by the model—to model training companies. That's where they actually make their money. In comparison, model adulteration is relatively minor.

The cost for ordinary users to discern this is too high. Some relay stations even sell you the real thing initially and then start adulterating later.

Many friends ask me how to screen for reliable relay stations. First, you must clarify which upstream source the channel is reverse-engineered from. Generally, the highest quality is the Claude service provided by Amazon AWS, followed by services reverse-engineered from Claude's official channels, and then channels with good industry reputation like Antigravity. If they can't clearly explain where it's extracted from, that's a red flag.

Different channels affect the output results. Those reverse-engineered from AWS and official channels will be very clean. Also, pay close attention to the cache hit rate. A simple example is: when you have three conversations with the model, can it read stored information from previous text? A high hit rate saves you money.

You can further test some typical features of official channels, such as native Web Search and multimodal image understanding capabilities. Native ones can do both. Send the same prompt simultaneously to the official model and the relay station, and compare the output quality, answer depth, and reasoning ability. If the relay station's answers are noticeably shallower and faster, it's a basic judgment that the version has been downgraded. You can also use third-party evaluation software.

In early April, Claude officially adjusted quotas, making the entire supply very tight, which had a huge impact on relay stations. That period was tough. I could only ask my technicians to build more different channels and also introduce some credible external channels to meet client demand.

This experience also helped me understand a counterintuitive way to discern reliability: if the model is under widespread risk control across the network and your pool dries up, clients are actually quite understanding. But if your pool doesn't dry up, clients will instead realize you've adulterated it. So, if you're severely blocked and have no quotas left, it actually indicates you're genuine.

In the C-end relay station business, the methods used to fight price wars and snatch clients are fierce. They suddenly post their own ads in your client groups, stage fake arguments, and more extreme ones will attack your website to make it go down.

The longest anyone has been in this industry is probably two or three years. The first two years, from what I know, were relatively easy, with some people getting rich overnight. Now it's become more difficult, partly because more people are doing it, and partly because risk control has improved.

So this year, many people just want to make a quick buck.

The most outrageous person I've seen was selling at an original price of 1.8, discounted to 0.8, and still claiming it was genuine. Users realized something was wrong by evening, argued with him, and he had no choice but to run away. There are also friends who deal mostly with individual clients, finding it increasingly troublesome, thinking more about pulling out in a couple of months.

"Notably, None from the United States"

I've been following this industry for almost a year and have joined different circles, but I've actually been operating a relay station for less than two months, all told.

I have clients both domestically and internationally, with about 80% in China. I basically only handle Claude, and incidentally provide GPT's 5.4 and 5.3. Clients coming for these two have relatively high requirements for model performance and output, many being pure development and R&D teams.

Early on, due to the popularity of OpenClaw, a batch of novice users very interested in Agents came. But later, due to Claude's various costs and daily usage habits, these users gradually disappeared. Those who remained till the end are still developers, R&D teams, and some B2B enterprises.

I also do SEO and GEO, so many overseas developers come seeking orders—from Europe, the Third World, Latin America, South Asia—notably, none from the United States.

There was even a white owner of a surf shop in Bali who curiously asked, "How do you sell this?" He also complained about how expensive Claude is. The whole world knows Claude is great to use, and the whole world also complains about Claude's price. Claude is an excellent model for people globally, but not everyone can afford it, except for enterprise-level users or companies that can provide a stable token supply.

What moved me the most were some clients from Third World countries, like India and Iran.

According to their monthly salary levels, Claude's price is a significant burden for them, forcing them to seek cheaper channels.

They directly ask if there's Sonnet. The price difference between Sonnet and Opus might be 2 to 3 times. They naturally focus on more cost-effective channels, and even on that basis, still use relatively cheaper models.

Some time ago, there was an Iranian client. Due to war-related blockades, they desperately needed smart AI to assist with many tasks, even risking government arrest. When communicating with them, I felt the war wasn't so far away from me.

Previously, we imagined doing overseas business meant dealing with relatively developed countries. But in fact, many Third World clients also face these troubles, also want to use AI, and are forced to turn to relay stations.

No one knows which group of users Anthropic truly wants to prevent with this KYC move—relay stations, distillation attack labs, or users from restricted regions. I think it's all of them.

But my feelings about this are complex. On one hand, it blocks the path for many ordinary users. Those who truly need Claude can only turn to relay stations, which is somewhat good for the relay station business. On the other hand, ordinary users find it hard to locate reliable channels, and the market will be disrupted by relay stations purely selling fake products.

Honestly, this isn't a good thing. People who can afford to keep a Claude subscription have high demands for stability. Spending time costs to also discern if there's adulteration might make everyone very disappointed with the whole relay station thing.

Reverse Relay Station

I don't consider myself a typical relay station operator.

The existence of relay stations is primarily due to information asymmetry. As long as there are clients, both domestic and international, who need cheaper APIs, this business will continue to exist.

I have a bold theory: if American models remain strong, relay stations will continue to exist because they will simply be expensive.

The USD exchange rate is there. Training the same model might cost $10 million in the US, but only 10 million RMB in China. This price gap will always exist.

But if Chinese models truly rise, becoming strong enough to beat American models, then relay stations will really have no place to play, because the acceptance of Chinese models would certainly be stronger than American ones.

In other words, the prerequisite for the long-term existence of this business is precisely the continued dominance of American models.

From the business itself, it has low barriers to entry—anyone can enter—yet a high ceiling. It can even be said to have a "mutually beneficial" relationship with the platform, because the more the platform blocks, the less likely relay stations are to disappear.

But I am overall pessimistic about AI.

Previously, I imagined that after the AI era arrived, many tedious physical labor tasks, basic mental labor, and dangerous jobs would be solved by machine intelligence, giving people time to focus on the spiritual world and humanity itself.

But later I discovered that even though many services already have regional pricing—for example, purchasing Claude models in Africa only costs 60%—that 60% is still a very expensive expense for locals.

In these regions, people who use AI are like having wings, while those who don't are walking on the ground. This speed will only widen the wealth gap further, concentrating more wealth in the hands of the already wealthy.

Although open-source models are advancing, those who can truly deploy them on their own devices for daily life services are still a minority.

Large models in the future might be an era that is both open-source and封锁 (blocked/restricted). I just want to try my best to allow everyone to equally use better intelligence. This is a very serious topic.

What I want to do long-term now is to take some open-source models or relatively cost-effective Chinese models overseas, directly providing them to some Latin American countries, less developed European countries, and further down to South Asia and Southeast Asia.

AI, in principle, should be something everyone can enjoy. From a business perspective, I will continue operating the relay station. But I tend to, outside of this business, do some truly open-source things. This is what I can see and what I can do.