Claude's Request for Identity Verification Prompts Reflection from a Relay Operator

Author | Kaori

Mo is the owner of a Claude intermediary station. In less than two months, she has served a group of clients, both domestically and internationally, who have high performance requirements for the Claude model. 80% of these clients are domestic, with the rest scattered across Europe, Latin America, South Asia, and the Third World, but "curiously, none in the United States."

She has also shared observations about this business on her website and social accounts, allowing some users who were originally unaware of intermediary stations to learn about the operation of this gray-area business model for the first time. The intermediary station plays a central role that requires constant adjustment among platform risk control, technical capabilities, and customer trust.

Mo describes herself as not a typical intermediary station owner. She does not engage in price wars but instead provides high-quality services only to B-end customers. She has a bold theory that for this business to exist in the long run, the U.S. model must dominate the market for an extended period. However, she is not optimistic about AI as a whole. Even with a region-based 40% discount, quality AI resources remain a heavy burden for many Third World countries.

Yesterday, news of Claude's introduction of passport KYC verification was released. Initially, Mo believed that this matter would not have a significant impact on the intermediary station industry. However, just one day later, the chilling effect of KYC was felt. What concerns her more is that this new regulation may push more ordinary users toward intermediary stations. Since the cost of discerning reliable channels is too high for ordinary users, this may not necessarily be a good thing.

Here is Mo's account:

Yesterday, news broke that Claude would be implementing passport KYC verification. Many people thought that intermediary stations would be completely out of business. However, Claude's current KYC is randomly triggered, so I thought my business would not be significantly affected for the time being.

The next day, I found that this matter had a negative impact on our business. The overall cost of number pools will increase, and channel stability will also be somewhat affected.

Whether it is hoarding numbers or finding KYC suppliers, the entire intermediary station business has many coping measures. Of course, this is only for now, and it is uncertain whether it will eventually be applied to all users.

Claude's decision to implement KYC is its own choice. Currently, the entire AI industry is in its growth phase. Other companies have not faced such high demand or other pressures like Claude, so other models have not implemented KYC yet. However, the future is uncertain.

The Intermediary Station Business

The intermediary station business is all about buying low and selling high, extracting account limits, selling them out within a certain payback period, and then making a profit.

The barrier to entry in this field is really low; as long as you have one or two clients on hand, you can get started.

I once met someone who charged a high price for their services, but when I visited their website, it had been running for six months and was still using HTTP. With a low defense rate against attacks, customer data was vulnerable to leaks.

Moreover, not everyone chooses to conduct this business cleanly.

One type involves packaging user call records. Some intermediary stations, in order to boost their consumer-facing metrics, offer seemingly unbelievably low prices and package backend databases—such as the input you provide to them via prompts and the content returned by models—to sell to model training companies. They actually make money from this transaction. Comparatively, model adulteration can be considered mild.

The cost of distinguishing for ordinary users is too high, and some intermediary stations initially provide genuine data but later adulterate it.

Many friends ask me how to select reliable intermediary stations. First and foremost, you need to clarify which upstream channel this service is reversely derived from. Generally, the highest quality comes from the Claude service provided by Amazon AWS, followed by services reverse-engineered from Claude's official source, and then channels with a good industry reputation like Antigravity. If they can't clearly explain the origin of the service, that should raise a red flag.

Different channels will influence the output results. Those reverse-engineered from AWS and official channels will be very clean. Additionally, you must pay attention to cache hit rates. For example, when you interact with a model three times, can it retrieve information from previous text, thus saving you money with a high hit rate.

Furthermore, you can test some typical features of official channels, such as native Web Search and multi-modal image understanding capabilities. The official ones can perform both tasks. By sending the same prompt to both the official model and the intermediary station simultaneously, you can compare output quality, response depth, and reasoning ability. If the station's response is noticeably shallower and quicker, you can conclude that the version has been downgraded. You can also use third-party evaluation software.

In early April, Claude officially adjusted its quota, causing a severe supply shortage, which greatly impacted intermediary stations. It was a tough period, and I had to develop various channels using my technology and introduce some trustworthy external channels to meet customer demand.

This experience also taught me an unconventional way of distinguishing. If the model exercises full-network risk control and your pool dries up, customers are generally quite understanding. However, if your pool hasn't dried up, customers will instead notice that you're adulterating the data. So, if you're heavily restricted, out of quota, it actually indicates authenticity.

In the B2C sector of this industry, the tactic of fighting for customers through price wars is very intense. Suddenly releasing your own advertisement to your customer base, engaging in a staged public quarrel, or even resorting to extreme measures such as launching cyber attacks to take down your website.

The longest-standing players in this industry have probably been around for only about two to three years. As far as I know, the first two years were relatively easy, with some experiencing overnight success. Now, it has become more challenging due to increased competition and enhanced risk management.

Therefore, this year, many people are just looking to make a quick buck.

The most outrageous person I've encountered was selling an item originally priced at 1.8, offering a discount to 0.8, and claiming it was genuine. However, when users discovered issues with the product that evening and confronted him, he disappeared. I also have friends who cater to retail customers and find it increasingly troublesome; their main thought now is to exit the market in a couple of months.

“Except there is no America”

I have been following this industry for nearly a year, participating in various circles, but in reality, engaging in the actual transfer station business adds up to less than two months.

Both domestic and international clients exist, with about 80% coming from domestic sources. Primarily, I only deal with Claude, occasionally providing GPT 5.4 and 5.3 as well. Clients who come based on these two models generally have high expectations for model performance and accuracy; many of them are pure development and research teams.

During the early days, due to the popularity of OpenClaw, a wave of newbie users interested in Agent arrived. However, later, due to various costs associated with Claude and daily usage habits, this user group gradually disappeared. The ones who remained till the end were developers, research teams, and some B2B enterprises.

I also handle SEO and GEO, so I receive many inquiries from overseas developers. I have inquiries from Europe, the Third World, Latin America, and South Asia, but notably, none from the United States.

There was even a surf shop owner from Bali who came over out of curiosity and asked, “How do you sell this?” He was a Caucasian and was also puzzled by why Claude was priced so high. The whole world knows that Claude is very user-friendly, and the whole world is also troubled by Claude's price. Claude is a great model for people worldwide, but not everyone can afford it, except for enterprise-level companies or those that can provide stable Token supply.

What touched me the most was some clients from Third World countries, such as India and Iran.

Based on their monthly salary levels, the price of Claude is a significant burden for them, forcing them to seek cheaper channels.

They directly ask if there is Sonnet available. The price difference between Sonnet and Opus could be 2 to 3 times, and they naturally gravitate towards more cost-effective channels, even looking for relatively inexpensive models on top of that.

A while back, we had a client from Iran who was experiencing blockades due to war and was in dire need of AI to assist with many tasks, even at the risk of government arrest. While communicating with them, I felt like the war wasn't that far away from me.

Previously, we thought of doing business overseas mainly with relatively developed countries. However, the reality is that many third-world clients also face these struggles and want to use AI, forcing us to turn to a transit hub.

No one knows which wave of users Anthropic's KYC this time is really trying to prevent—transit hubs, distillation attack labs, or users from restricted regions. I believe it encompasses all of them.

My sentiments about this matter are quite complex. On the one hand, it closes off the path for many regular users, and those who truly need Claude can only turn to a transit hub, which is somewhat beneficial for the transit hub business. On the other hand, regular users find it challenging to find a reliable channel, as the market is disrupted by transit hubs that only sell counterfeit goods.

To be honest, this is not a good thing. Those who have the ability to maintain a Claude subscription have high stability requirements. Spending time and effort to distinguish whether there is any adulteration may leave everyone very disappointed with the transit hub concept.

Reverse Transit Hub

I don't think of myself as a very typical transit hub operator.

The existence of transit hubs is primarily due to information asymmetry. As long as there are customers always in demand for cheap APIs, regardless of domestic or international, this business will continue to exist.

I have a radical theory: if the U.S. model continues to dominate, transit hubs will persist because it will be very costly.

With the USD exchange rate as it is, training a model in the U.S. may cost $10 million, while in China, it might only require 10 million RMB, and this price difference will always remain.

However, if the Chinese model truly emerges and becomes strong enough to surpass the U.S. model, then transit hubs will truly have no place because the Chinese model's acceptance will undoubtedly be stronger than that of the U.S.

In other words, the premise for the long-term existence of this business is precisely the long-term dominance of the U.S. model.

From a business perspective, it has a low barrier to entry; anyone can join, and yet the ceiling is high. One could even say it mutually benefits platforms because the more they tighten restrictions, the less likely transit hubs will disappear.

But I am overall not optimistic about AI.

Previously, when envisioning the AI era, I thought that many tedious physical tasks, basic mental labor, and dangerous jobs would be solved by machine intelligence, allowing people to focus on the spiritual world and humanity.

However, it was later discovered that even though many services were already priced based on region, such as a 40% discount for purchasing the Claude model in Africa, this 40% discount was still a significant expense for the local population.

In these regions, those who have AI are like they have grown wings, while those without AI are merely walking on the ground. This discrepancy in speed will only widen the wealth gap further, concentrating more wealth in the hands of the already wealthy.

Although open-source models are being promoted, only a few can actually deploy them on their own devices for everyday use.

In the future, large models may exist in a time that is both open-source and closed. I want to make every effort to ensure that everyone can equally access better intelligence; this is a very serious issue.

What I have been thinking about for a long time is taking some open-source models or models that offer relatively high cost-effectiveness in China and introducing them overseas, directly providing them to some Latin American countries, underdeveloped European countries, and then extending to South Asia and Southeast Asia.

AI, in theory, should be accessible to everyone. From a business perspective, I will continue to act as a middleman, but I am inclined to engage in some truly open initiatives beyond this business. This is something I can see and do.