Claude's KYC Implementation Sparks Debate in the Proxy Station Industry

Original author: Kaori

Mo is the administrator of a Claude relay station. In the past two months, she has served a group of clients with high requirements for Claude model performance, 80% of whom are in China, while the rest are scattered across Europe, Latin America, South Asia, and the developing world—none in the United States.

She has also shared insights about this business on her website and social media accounts, introducing some users who were previously unaware of money mules to the way this gray area operates. Money mules serve as an intermediary that continuously balances platform risk controls, technical capabilities, and customer trust.

Mo describes herself as not a typical transit station operator—she doesn’t engage in price wars, instead focusing on delivering high-quality service to B2B clients. She holds a bold view: the long-term viability of this business depends on the U.S. model remaining dominant. Yet she remains pessimistic about AI overall; even with region-based discounts of up to 40% off, premium AI resources remain a heavy burden for many people in the Global South.

Yesterday, when news broke that Claude had launched passport KYC verification, Mo initially judged that the impact on the intermediary industry would be limited. But just one day later, the chill of KYC regulations had already spread. What concerned her more was that this new rule might push more ordinary users toward intermediary services—yet the cost for ordinary users to identify reliable channels is too high, which isn’t necessarily a good thing.

Here is Mo's account:

When news spread yesterday that Claude was undergoing passport KYC verification, many assumed the transit service had been completely shut down. However, Claude’s current KYC is randomly triggered, and I initially believed my business would not be significantly affected for now.

The next day, we still found that this incident had a negative impact on our business, leading to increased overall costs for the account pool and some impact on channel stability.

Whether hoarding accounts or finding KYC providers, there are many countermeasures for this intermediary business. Of course, this is just the current situation—it’s uncertain whether it will eventually be applied to all users.

Claude’s decision to implement KYC is its own choice. Currently, the AI industry is still in its growth phase, and other companies have not yet faced the same level of demand or pressure as Claude, which is why other models have not implemented KYC. However, the situation may change in the future.

The arbitrage business involves buying low and selling high—extracting account limits and selling them off within a specific payback period to generate profit afterward.

The barrier to entry in this industry is very low—all you need is one or two clients to get started.

I once encountered someone whose services were quite expensive, but when I visited their website, I found it was still using HTTP after six months. Their defense against attacks was weak, making customer data highly vulnerable to leaks.

Moreover, conducting this business cleanly isn't an option for everyone.

One type involves bundling users’ interaction logs; some intermediaries inflate their consumer-side volume by selling backend data—such as your prompts and the model’s responses—to model training companies at seemingly absurdly low prices. Their real profit comes from this practice. In comparison, model tampering is relatively minor.

The cost of verification for ordinary users is too high, and some intermediaries initially sell genuine products but later mix in counterfeits.

Many people ask me how to identify reliable relay services. First, you must clarify which upstream source this channel is reverse-engineered from. Generally, the highest quality is the Claude service provided by Amazon AWS, followed by services reverse-engineered directly from Claude’s official platform, and then reputable channels like Antigravity. If they can’t clearly explain where it’s sourced from, that’s a red flag.

Different channels can affect the output results; those reverse-engineered from AWS and official channels will be much cleaner. Also, pay close attention to cache hit rates—for example, when you converse with the model three times, can it retrieve previously stored information from earlier text? A higher hit rate saves you money.

You can further test typical characteristics of official channels, such as native web search and multimodal image understanding. Native models can perform both tasks. Send the same prompt simultaneously to the official model and the intermediary, then compare output quality, depth of response, and reasoning capability. If the intermediary’s response is noticeably shallower and faster, it is likely that the version has been downgraded. You can also leverage third-party evaluation tools.

In early April, Claude officially adjusted its quota, causing the entire supply to become extremely tight and significantly impacting our intermediary services. During that period, it was very challenging, so I had my technical team create multiple alternative channels and also integrate some trusted external channels to meet customer demand.

This experience also taught me a counterintuitive way to identify authenticity: if the model triggers system-wide risk controls and your pool dries up, customers are generally understanding. But if your pool remains full, customers may suspect you’re manipulating the system. So, if you’re heavily restricted and have no quota left, it actually suggests you’re genuine.

On the C-end, the competition for intermediary services is extremely fierce, with players aggressively fighting over customers through price wars. Some suddenly flood your customer groups with their own ads, stage fake arguments to mislead users, and even more extreme cases involve launching attacks on your website to cause it to go down.

The longest anyone has been in this industry is probably two or three years. In the first two years, as far as I know, it was relatively easier, and there were people who became rich overnight. Now it’s much harder—on one hand, more people are entering, and on the other, risk management has improved.

So this year, many people wanted to make a quick profit.

I’ve met the most outrageous person—originally priced at 1.8 yuan, now discounted to 0.8, yet still claiming it’s genuine. The user noticed something was off by evening and confronted him, but he just vanished. I also know friends who deal with many individual customers; they’re increasingly finding it troublesome and are planning to quit in two months.

I’ve been following this field closely for nearly a year and have joined various communities, but I’ve only been actively operating as a intermediary for less than two months.

We serve both domestic and international clients, with domestic clients accounting for approximately 80%. We primarily offer Claude, along with GPT-5.4 and GPT-5.3 as supplementary options. Clients attracted by these models have high requirements for performance and reliability, and many are dedicated development and R&D teams.

Initially, OpenClaw's popularity attracted a wave of novice users highly interested in agents. However, due to Claude’s various costs and everyday usage habits, these users gradually disappeared. Those who remained were primarily developers, research teams, and some B2B enterprises.

I also do SEO and GEO, so many overseas developers come to place orders—from Europe, the developing world, Latin America, and South Asia—but none from the United States.

Even the owner of a surf shop in Bali came over curiously asking, “How much does this cost?” He’s a white man who also struggles with how expensive Claude is. Everyone around the world knows Claude is incredibly useful, and everyone around the world struggles with its price. Claude is an excellent model for people everywhere, but not everyone can afford it—only enterprises or companies with a stable token supply can.

What moved me the most were customers from some developing countries, such as India and Iran.

Given their monthly salary levels, Claude’s pricing is a significant burden, forcing them to seek cheaper alternatives.

They will directly ask if Sonnet is available; the price difference between Sonnet and Opus may be two to three times, so they naturally gravitate toward more cost-effective options and may even opt for relatively cheaper models on that basis.

A while ago, I had an Iranian client who was blocked due to war and desperately needed smart AI to assist with many tasks, even risking arrest by their government. Communicating with them made me feel like war wasn’t as far away as I thought.

Previously, we imagined doing business overseas would focus on relatively developed countries, but in fact, many clients from developing nations also face these challenges and want to use AI, forcing us to turn to intermediaries.

No one knows which group of users Anthropic is truly trying to prevent with this KYC—intermediaries, distillation attack labs, or users in restricted regions—I think it could be all of them.

But my feelings about this are complicated: on one hand, it blocks the path for many ordinary users, forcing those who truly need Claude to turn to intermediaries, which is somewhat beneficial for intermediary services. On the other hand, ordinary users struggle to find reliable channels, and the market is disrupted by intermediaries selling nothing but fake services.

To be honest, this isn't a good thing. Those who can afford to maintain a Claude subscription consistently have high expectations for stability. Spending time and effort just to determine whether something is authentic may leave people deeply disappointed in the intermediary service.

I don’t consider myself a typical transit station manager.

The existence of transit stations is primarily due to information asymmetry; as long as there are customers demanding cheaper APIs, whether domestic or international, this business will continue to exist.

I have a bold claim: if the U.S. model remains strong, intermediaries will continue to exist because they will always be expensive.

The exchange rate for the US dollar is what it is; training the same model might cost $10 million in the United States but only 10 million RMB in China—this price difference will always exist.

But if the Chinese model truly rises and becomes strong enough to surpass the American model, then the intermediary model would truly be unsustainable, because the acceptance of Chinese models would certainly be stronger than that of American ones.

In other words, the long-term viability of this business depends precisely on the U.S. model maintaining its long-term dominance.

From a business perspective, it has low entry barriers and is open to anyone, with a very high ceiling—so high that it can be said to mutually benefit the platform, because the more the platform restricts it, the more the intermediary station persists.

But I’m generally not optimistic about AI.

Previously, I envisioned that with the arrival of the AI era, many tedious physical tasks, basic mental labor, and dangerous jobs would be handled by machine intelligence, allowing people to focus on their inner world and themselves.

However, it was later found that even though many services are already priced by region—for example, purchasing the Claude model in Africa costs only 40% less—it is still an expensive expense for locals.

In these regions, people who use AI are like they’ve grown wings, while those who don’t use AI are still walking on the ground. This speed gap will only widen the wealth divide, concentrating even more wealth in the hands of those who are already rich.

Although open-source models are advancing, only a few can actually be deployed on personal devices to serve daily life.

The era of large models may soon be one where they are both open-source and restricted. I want to ensure that everyone can equally access better intelligence—it’s a serious issue.

What I’ve long wanted to do is take open-source models or relatively cost-effective models from China and take them global, directly targeting Latin American countries, less developed European nations, and further down to South and Southeast Asia.

AI, in principle, should be accessible to everyone. From a business perspective, I’ll continue operating as a middleman, but I’m inclined to dedicate efforts beyond this business to truly open-source initiatives—this is something I can see and do.