Here's one of the three mentioned reasons why they're cheap:
> Swapping models and inflating tokens. Because users’ inputs and model outputs are mediated through a proxy, users cannot verify which model their request was actually routed to. A user selects Opus 4.7, but the proxy can silently route to Sonnet, Haiku, or, in the worst case, GLM or Qwen, and fraudulently relabel the output. In a recent paper from Germany’s CISPA Helmholtz Center for Information Security (which cited my article last year on grey market!), researchers audited 17 API proxies and found widespread model swapping–API proxy access to “Gemini-2.5” achieved only 37.00% on a medical benchmark, a staggering drop from the 83.82% performance of the official API. On the user end, the tell only comes on complex tasks, when the output feels off (often referred to as 降智, or “dumbed-down”), but there is no clean way to prove it. Numerous public records highlight concerns that certain API proxies have noticeably compromised model performance. These proxies are suspected of “diluting” (掺水) services by substituting premium frontier models with inferior tiers.
So no, those cheap tokens won't necessarily be Claude.
Odd that they'd risk getting screwed over like that, when DeepSeek v4 Pro is pretty okay nowadays for quite a few tasks. I guess it's a bit like OpenRouter, where I get to try out all sorts of models with relatively few hassles (though nobody will give me a discount), but I have to acknowledge that some providers will straight up quantize the models so far that they're borderline unusable.
Interesting article, though nothing exactly new or surprising about KYC and anti-spam methods based on phone numbers and credit cards being fundamentally flawed and producing gray-market solutions.
Still, personally i think there's one piece missing in the article. Why would it be OK to restrict chinese users from using american models? I mean, personally i'm strongly anti-AI and i believe all AI companies need to die because they enhance the worst humanity has to offer. However, if AI is going to be legal, how can it be ethical to discriminate based on one's country? Especially if said country (China) is the one refining 90% of the minerals and rare earths the US uses to produce its computers.
Well, Dario Amodei used to work in Baidu's SVAIL lab, and he certainly noticed the shady side of Chinese business practices and how he already has a prejudice that the Chinese are unfaithful. As a Chinese myself I don't really want to blame him because I know how's that working out first hand too.
In business nobody can be expected to be faithful. Funny enough I've never been screwed over by chinese but a plenty of times by Europeans and americans. Despite the fact I deal with Chinese more.
Dario Amodei himself has a great incentive to lie (FUD and denying the others the AI R&D, also his nationalism is well known from his geopolitical rants). In particular, he claimed without presenting any evidence that DeepSeek is distilling Claude, but if you know anything about those two models you know they have absolutely nothing in common (especially the CoT).
One of these statements is true: 1. DS have found a secret way to distill models while not making them write similarly to the source, and are hiding it. 2. Dario Amodei is lying through his teeth. Considering that all previous known attempts at making reward models from other model outputs were easily visible (Gemini 2 experimentals from Claude and GLM 4.6+ from Gemini 2.5/3.0, both took a lot of character from source models and GLM even copied Google's prompt injections), DeepSeek are remarkably open, and DeepSeek v4 is pretty consistent in character with 3.1+, suggesting similar post-training, statement 2 seems to be way more likely.
Previously Sam Altman claimed the same about DS v3, also without any evidence and the models were nothing alike as well.
Interesting article, one thing I do not get, why go through all these risks and trade-offs just to be able to use Anthropics models? I've tried Claude Opus/Sonnet/Haiku, they are good. But are they really THAT good that no other model even comes close for the clients use case? We have Deepseek V4, Qwen 3.6, GLM-5.1, Kimi k2.6 even Chatgpt 5.5! Why are the so stubborn to use Claude?
Interesting article! You can basically everything cheap on Taobao, wonder if they use the same principles, i am talking adobe cloud subscriptions etc. Also.. netease music allows hq downloads of the whole library for like 1usd a month. They have everything, even super niche stuff, not sure how that works… Anyway, it is interesting how Claude’s protections work. Chinese users have been dealing with these types of restrictions for a decade or so now and are super savvy in circumvention, up until now its not really a cat and mouse game
Here's one of the three mentioned reasons why they're cheap:
> Swapping models and inflating tokens. Because users’ inputs and model outputs are mediated through a proxy, users cannot verify which model their request was actually routed to. A user selects Opus 4.7, but the proxy can silently route to Sonnet, Haiku, or, in the worst case, GLM or Qwen, and fraudulently relabel the output. In a recent paper from Germany’s CISPA Helmholtz Center for Information Security (which cited my article last year on grey market!), researchers audited 17 API proxies and found widespread model swapping–API proxy access to “Gemini-2.5” achieved only 37.00% on a medical benchmark, a staggering drop from the 83.82% performance of the official API. On the user end, the tell only comes on complex tasks, when the output feels off (often referred to as 降智, or “dumbed-down”), but there is no clean way to prove it. Numerous public records highlight concerns that certain API proxies have noticeably compromised model performance. These proxies are suspected of “diluting” (掺水) services by substituting premium frontier models with inferior tiers.
So no, those cheap tokens won't necessarily be Claude.
Odd that they'd risk getting screwed over like that, when DeepSeek v4 Pro is pretty okay nowadays for quite a few tasks. I guess it's a bit like OpenRouter, where I get to try out all sorts of models with relatively few hassles (though nobody will give me a discount), but I have to acknowledge that some providers will straight up quantize the models so far that they're borderline unusable.
Knock-off tokens from a replica router, that's hilarious. "If you look closely, the logo says Clod."
Interesting article, though nothing exactly new or surprising about KYC and anti-spam methods based on phone numbers and credit cards being fundamentally flawed and producing gray-market solutions.
Still, personally i think there's one piece missing in the article. Why would it be OK to restrict chinese users from using american models? I mean, personally i'm strongly anti-AI and i believe all AI companies need to die because they enhance the worst humanity has to offer. However, if AI is going to be legal, how can it be ethical to discriminate based on one's country? Especially if said country (China) is the one refining 90% of the minerals and rare earths the US uses to produce its computers.
Nvidia chips are also legal, yet restricted. No need to invoke ethics when you have power.
Well, Dario Amodei used to work in Baidu's SVAIL lab, and he certainly noticed the shady side of Chinese business practices and how he already has a prejudice that the Chinese are unfaithful. As a Chinese myself I don't really want to blame him because I know how's that working out first hand too.
In business nobody can be expected to be faithful. Funny enough I've never been screwed over by chinese but a plenty of times by Europeans and americans. Despite the fact I deal with Chinese more.
Same experience here.
Dario Amodei himself has a great incentive to lie (FUD and denying the others the AI R&D, also his nationalism is well known from his geopolitical rants). In particular, he claimed without presenting any evidence that DeepSeek is distilling Claude, but if you know anything about those two models you know they have absolutely nothing in common (especially the CoT).
One of these statements is true: 1. DS have found a secret way to distill models while not making them write similarly to the source, and are hiding it. 2. Dario Amodei is lying through his teeth. Considering that all previous known attempts at making reward models from other model outputs were easily visible (Gemini 2 experimentals from Claude and GLM 4.6+ from Gemini 2.5/3.0, both took a lot of character from source models and GLM even copied Google's prompt injections), DeepSeek are remarkably open, and DeepSeek v4 is pretty consistent in character with 3.1+, suggesting similar post-training, statement 2 seems to be way more likely.
Previously Sam Altman claimed the same about DS v3, also without any evidence and the models were nothing alike as well.
Interesting article, one thing I do not get, why go through all these risks and trade-offs just to be able to use Anthropics models? I've tried Claude Opus/Sonnet/Haiku, they are good. But are they really THAT good that no other model even comes close for the clients use case? We have Deepseek V4, Qwen 3.6, GLM-5.1, Kimi k2.6 even Chatgpt 5.5! Why are the so stubborn to use Claude?
Interesting article! You can basically everything cheap on Taobao, wonder if they use the same principles, i am talking adobe cloud subscriptions etc. Also.. netease music allows hq downloads of the whole library for like 1usd a month. They have everything, even super niche stuff, not sure how that works… Anyway, it is interesting how Claude’s protections work. Chinese users have been dealing with these types of restrictions for a decade or so now and are super savvy in circumvention, up until now its not really a cat and mouse game