Introduction

Mo is the manager of a Claude proxy station. In the past two months, she has served a group of clients with high demands for Claude’s model performance, with 80% based in China and the rest scattered across Europe, Latin America, South Asia, and the developing world, notably excluding the United States.

She has shared her observations about this business on her website and social media, introducing users to the workings of this gray area. Proxy stations play a mediating role, calibrating between platform risk control, technical capabilities, and customer trust.

Mo describes herself as not a typical proxy station manager; she does not engage in price wars and focuses on providing quality service to B2B clients. She holds a controversial view that the long-term existence of this business relies on the continued dominance of American models.

However, she is not optimistic about AI overall. Even with a 40% discount for many developing countries, high-quality AI resources remain a heavy burden for many.

Impact of KYC Verification

Yesterday, the news about Claude’s passport KYC verification emerged. Initially, Mo thought this would not significantly impact the proxy station industry. However, just a day later, she realized that KYC had negative effects on their business, increasing overall costs and affecting channel stability.

Whether hoarding accounts or finding KYC suppliers, there are many coping strategies in the proxy station business. This is just the current situation; it is uncertain whether KYC will become widespread among all users.

Claude’s KYC is its own choice. The entire AI industry is still in an upward phase, and other companies do not face the same demand or pressure as Claude, which is why other models have not implemented KYC yet.

The Proxy Station Business

The proxy station business involves buying low and selling high, extracting account quotas and selling them within a certain payback period to make a profit.

The entry barrier for this industry is very low; anyone with one or two clients can start.

Mo recalls encountering someone who sold services at a high price, but their website was still HTTP after six months, making them vulnerable to attacks and risking client data leaks.

Not everyone chooses to run a clean operation. Some proxy stations package user call records and sell backend databases, including prompts and model responses, to training companies, profiting from this instead.

Ordinary users find it hard to discern reliable channels. Some proxy stations may initially provide genuine services but later compromise quality.

When asked how to filter reliable proxy stations, Mo suggests first clarifying the upstream source of the channel. Generally, the highest quality comes from Claude services provided by Amazon AWS, followed by services directly from Claude, and then reputable industry channels like Antigravity. If a provider cannot clarify their source, it raises a red flag.

Different channels affect output quality. Services from AWS and official channels are typically very clean. It’s also essential to monitor cache hit rates; for example, when conversing with the model multiple times, it should be able to reference previous texts, which saves costs.

Further, one can test typical features of official channels, such as native web search and multimodal image understanding. If a proxy station’s responses are noticeably shallower and quicker than the official model’s, it likely indicates a downgraded version. Third-party evaluation software can also assist in this assessment.

In early April, Claude adjusted quotas, tightening supply and significantly impacting proxy stations. During that period, Mo had to diversify her channels and introduce trustworthy external sources to meet client demands.

This experience led her to an unconventional discernment method: if the model faces widespread risk control and your pool runs dry, clients tend to be understanding. However, if your pool remains full, clients may suspect quality issues. Thus, being heavily restricted and out of quota might ironically indicate genuine service.

In the C-end proxy station business, competition is fierce, with aggressive tactics to steal clients, including spamming ads in client groups or even attacking competitors’ websites.

The industry has only been around for two to three years, and the first two years were relatively easy, with some people becoming overnight successes. Now, it’s becoming more challenging due to increased competition and enhanced risk control.

Many people this year are looking to make quick money.

Mo recalls one outrageous case where a service priced at 1.8 was discounted to 0.8, claiming authenticity. Users later discovered discrepancies and the seller fled.

The Absence of American Clients

Having followed this industry for nearly a year and being involved in various circles, Mo has only been operating a proxy station for about two months.

Her clients are both domestic and international, with about 80% from China. She primarily offers Claude services, along with GPT-5.4 and 5.3. Clients coming for these models typically have high expectations for performance and capabilities, often being development teams.

Initially, there was a surge of novice users interested in Agent due to OpenClaw’s popularity, but this group gradually disappeared due to Claude’s costs and usage habits. Ultimately, the remaining clients are developers, R&D teams, and some B2B enterprises.

Mo also engages in SEO and GEO, attracting many overseas developers from Europe, the developing world, Latin America, and South Asia, but notably not from the United States.

She even had a surf shop owner from Bali inquire about pricing, expressing frustration over Claude’s high costs.

Globally, everyone knows Claude is effective, yet many are troubled by its price. While Claude is a great model for people worldwide, not everyone can afford it except for enterprise-level clients or those providing stable token supplies.

What resonates with Mo are clients from developing countries, such as India and Iran.

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

They often ask if there are alternatives like Sonnet, where the price difference compared to Opus can be two to three times. They naturally seek more cost-effective options, often needing to use relatively cheaper models.

Recently, an Iranian client expressed the urgent need for intelligent AI due to war-related restrictions, risking government arrest to access necessary tools. Conversations with them made the realities of war feel closer.

Previously, Mo thought that overseas business would primarily involve developed countries, but many clients from developing nations face similar struggles and turn to proxy stations for solutions.

No one knows the true intent behind Anthropic’s KYC implementation—whether it’s aimed at proxy stations, distilled attack labs, or users from restricted regions—but Mo feels conflicted. On one hand, it closes off access for ordinary users, pushing those who genuinely need Claude towards proxy stations, which could benefit the proxy business. On the other hand, ordinary users struggle to find reliable channels, leading to market disruption from dishonest proxy stations.

Honestly, this is not a good situation. Users who can maintain a Claude subscription have high stability requirements. The time spent discerning quality can lead to disappointment with proxy stations.

Reverse Proxy Stations

Mo does not consider herself a typical proxy station manager.

The existence of proxy stations is primarily due to information asymmetry. As long as there is a demand for affordable APIs, both domestically and internationally, this business will persist.

She holds a controversial view: if American models continue to thrive, proxy stations will persist because they remain expensive.

The dollar exchange rate is a factor; training a model in the U.S. might cost $10 million, while in China, it could be only 10 million RMB, maintaining a permanent price gap.

However, if Chinese models rise to a level that can surpass American models, the proxy station business could collapse, as acceptance of Chinese models would likely be higher.

In essence, the long-term existence of this business hinges on the sustained dominance of American models.

From a business perspective, the entry barrier is low, anyone can participate, and the ceiling is high. It can be said that the platform and proxy stations mutually benefit each other; the more the platform restricts access, the more proxy stations will persist.

Yet, Mo remains generally pessimistic about AI.

She once envisioned an AI-driven future where tedious physical labor and basic cognitive tasks would be handled by intelligent machines, allowing humans to focus on more profound aspects of life.

However, she has realized that even with regional pricing adjustments, such as a 40% discount for Africa, the cost remains prohibitively high for locals.

In these regions, those who can access AI feel empowered, while those who cannot are left struggling. This disparity only exacerbates the wealth gap, concentrating more wealth in the hands of the affluent.

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

The future of large models may be a mix of open-source and restricted access. Mo aims to ensure that everyone can equally access better intelligence, a serious challenge.

Currently, her long-term goal is to promote open-source models or relatively cost-effective Chinese models to Latin American countries, less developed European nations, and further down to South Asia and Southeast Asia.

AI should ideally be accessible to all. While she will continue her proxy station business, she also intends to pursue genuinely open-source initiatives, which is something she can envision and work towards.