The consequences of locking intelligence away: an introduction to Claude relays in China
There has been recent discourse floating around on Hacker News about Chinese API relay stations that use every Western VC-subsidized channel of cheap tokens (think Claude/ChatGPT subscriptions, AWS/Azure credits, Kiro, Google Antigravity, etc.) to resell as APIs to the domestic Chinese market.
This is true, as a Chinese citizen that has been seeing an uptick of this trend since mid 2024, but especially since 2025. When I go on Taobao (China’s Amazon) and search for keywords, there would be dozens of relay services selling for around 1/5th to 1/10th the price of official western APIs.
Indeed, many of these relays may have cheaper models disguised as genuine western ones, so the Chinese tech community has entire forums such as linux.do that serves primarily as a way for people to discuss and rate relay services based on their price, quality, and availability, as well as websites such as hvoy.ai that uses a variety of automated testing suites to benchmark the quality of relay providers. There are even free relays that exists primarily as a gentleman’s handshake data collection method between relay operators and users—some in China have speculated that one of the most popular ones is operated by Zhipu, which might explain their model’s apparent claude-ness.
This is also why Chinese LLMs are far cheaper on average: they both have to compete with each other, but also with grey market relays serving foreign models at aggressively subsidized rates in the domestic market. Even GLM-5.2′s $4.2/million out price is having a hard time competing with Opus 4.8 at a $7.5/million out price in the most reputable API relay services.
In retrospect, this is the natural endgame of current era China AI bans.
When people in the West hear about Anthropic accusing Chinese firms of distilling their models, they map them onto the typical cases of corporate espionage: innovative U.S. companies create the frontier of intelligence, then the Chinese copy them in massive government-subsidized campaigns. However, it’s far from the case. There simply is sufficient market incentive for both tech enterprises as well as individual developers to onboard third-party relays selling Claude Opus 4.8 in 1/10th the price.
I note my conflict of interest: I’m naturally frustrated at how I’m unable to access Anthropic’s services legitimately as they have been cracking down on users with VPNs, so this channel is the lifeline I get to interacting with SOTA models. however, I’m also highly aware of the numerous data risks that surrounds this grey market (which is what caused relays to be named in Chinese state media as an area for concern—interestingly, China has always been lax to programmers and STEM students in terms of the GFW, so “foreign models tells that Taiwan is a country” and such are not named).
Perhaps post-Mythos and post-GPT-5.6 era KYC would reduce this phenomenon, but there has already been workarounds brewing (examples being reverse engineering the APIs of third party agentic coding services), so we’ll have to see about their efficacy.
I’ll be happy to take questions from the comment section.
Why are the tokens so cheap in China? Who is subsidizing this and why?
What does “gentleman’s handshake data collection method between relay operators” mean?
OpenAI and Anthropic are subsidizing the subscriptions themselves, as price discrimination (enterprise customers will pay the full API price, but they also want to capture the individual subscription market). This also incentivizes people to use Codex / Claude Code rather than competitors like Cursor. Not sure about the other subsidized channels.
It’s a really substantial discount, probably over 10x if you fully use the weekly allowance of a Max 20x subscription.
I’m also guessing that like a gym membership, many of the subscriptions are under-used in practice. As someone whose work and hobbies don’t involve significant programming, I have a Claude Max subscription but I only come close to limits very rarely (the only times in recent memory was when I tried to use Fable on Max to do a sweep of questions that interested me and barely seemed on the cusp of doable with present-day AI, in the 3 days I had access to Fable[1]).
I got blocked/filtered a bunch of times on work-relevant tasks but I wanted to see if it had novel literary insights on Chiang, Kafka, etc. Burned a bunch of tokens on this.
Agree. I expect them to be more fully used in the future. As inference scaling improves, it means there will be more tasks that can be done with 100% of token limits but not 10% of token limits. Companies will probably adjust pricing strategy in response.
Hi! Your question relayes to the several ways for relay operators are incentivized to operate in 2026 that I understand.
On one hand, efficiently routing and utilizing each Claude 5x or 20x sub means gaining access to around $4000 worth of tokens with just $200 a month of subscription. OpenAI subsidizes it even more aggressively, reaching more than $5000. However, the cost of being banned is also high (partially due to surveillance built into Claude Code which flags the most popular third party relays). Some relay operators may open-source parts of their routing code on Github (on top of using open source relay builders like NewAPI) to serve enterprises that have privacy requirements. Many have even implemented OpenRouter-esque data protection and personal information redaction features for orgs to choose to implement for their employees.
But on the other hand, many programmers may not want to put money into services that might get shut down at any time; this is where “公益站” (free relays) come into play. These are typically advertised in tech forums, and are primarily focused on collecting out-of-distribution data to be distilled into models. That’s what I mean by a “gentleman’s handshake data collection method between relay operators”—the user implicitly understand that their data is being used for distillation purposes, and understands the risk well enough to justify using it in certain contexts.
Would you want to access the services legitimately if you can? It seems like if if it’s easy and socially acceptable to use the services 10x cheaper, the case for paying full-price is much weaker unless you aren’t price-sensitive at all.