Chinese resellers are offering Claude tokens at 70-90% below official Anthropic API prices. They achieve this by reselling capacity from pooled Claude Max accounts, payments fraud, and also reselling the model output & reasoning chains to various Chinese labs. They are subsidizing model access in exchange for user logs and reasoning traces, which they then sell as training data, allowing them to operate below cost.
Claude and ChatGPT are both blocked in China. You need to use a VPN to access either, and you can't pay with a Chinese bank card. So most people who want access to Claude buy access via a reseller. It's the easiest and cheapest way to access Anthropic models in China.
These resellers operate tens of thousands of bot accounts, which is also why Anthropic introduced identity verification, to slow down the onslaught of bots.
This is one reason why DeepSeek & GLM are priced so cheaply, they are competing with impossibly low token prices in China. They have to keep prices low, in order for people to use them.
Ok, but what about those shady sites that resell Windows education keys? They're certainly a "better experience" than buying legit keys, by virtue of being significantly cheaper. You aren't even really committing copyright infringement in the process, because Microsoft gives out windows isos for free, and the seller is really selling a random 25 character string, which can hardly be copyrighted.
>If there’s any bootlegging going on it’s Anthropic that’s doing the bootlegging but having mirrored the video etc sufficiently to beat copyright law.
I get the vague impression that this was written in a sarcastic way, but it has a straightforwardly true literal read because yes, this is what the free market is about and Anthropic will have to compete with the Chinese if they want a big share of the market. Chinese models are cheap and good; even without reselling Anthropic's services they're competitive. Which reading did you intend?
And, gotta say, the idea that the Chinese are better at selling US models than the Americans is hilarious. There might be an economic study here somewhere about just how anti-consumer and anti-progress their IP laws turned out to be. We've got an entire postindustrial revolution centred around who can ignore the most stupid laws.
What is the implication here? Are you warning that US corporations might start doing something shady, like scraping the internet at large scale for training data? Or mass-dowloading pirated copies of books, completely ignoring copyright?
I find it hard to imagine a future where US corporations have degraded to such a point.
No, he means that the US will close most of its domestic market to competition just like China has for decades, and the US may start subsidizing and dumping its goods everywhere
Most of the Chinese domestic market is open to foreign competition. The areas that are closed off are those that are politically sensitive: publishing (including social media) and banking.
As for dumping, Chinese goods generally sell at a markup abroad, which is the opposite of dumping. Chinese tokens cost more abroad. Chinese cars cost several times more in Western markets than in China.
In debt the first 5000 years Geaeber makes the case that pure “free market” trade has never really existed in “the west”. The closest to this ideal that’s ever happened was during the Islamic golden age enabled by religious prescriptions against usury.
>The closest to this ideal that’s ever happened was during the Islamic golden age enabled by religious prescriptions against usury.
How does are bans against consensual financial exchanges close to the "ideal" of the free market? It just sounds like you have an axe to grind about the financial system rather than describing free markets.
What makes this view more correct than say, "economies with marketing creates a dynamic where being competitive in production is secondary to marketing" and concluding that nothings a free market until we ban all advertising? After all, you can make a vaguely plausible argument about how marketing isn't really about the merits of the product, and therefore allowing it is antithetical to the free market or whatever
> What makes this view more correct than say, "economies with marketing creates a dynamic where being competitive in production is secondary to marketing" and concluding that nothings a free market until we ban all advertising? After all, you can make a vaguely plausible argument about how marketing isn't really about the merits of the product, and therefore allowing it is antithetical to the free market or whatever
Wait, so your pitch in favor of a debt-fueled market economy is that advertising is awesome and that we wouldn't want to "lose" being smothered in ads all the time?
Cause... sign me up for the non-financialized, non-mass-media-advertising-driven economy please and thank you. I'd even be ok with just nuking billboards and mass-media forms of ads and still allowing more direct forms of marketing, if we must compromise!
(I thought the disconnect between the efficiency of competition and the market as realized in modern economies was pretty well understood and taken for granted, but I guess we all find ways to justify the system we're profiting from... even if that means we have to claim we love the ad breaks)
AI was always going to be a race to the bottom and low margins. It’s why I’m extremely bearish on AI as an investment. It’s framed as some high margin business when it’s really going to end up like your toilet paper at Costco. You will use whatever is cheapest and gets the job done.
I used to think this.. but I think my opinion is changing. The reason is that the leaders likely will be able to accelerate faster.
So what you see is the market "stretching".. the bottom getting cheaper and the top end running away and getting more expensive. At some point the top end may be too valuable to even sell access to.
Glm 5.2 very much argues against that. Opus 4.8 level quality for cheap. That’s sufficient for most tasks, so if/when you do need SOTA models you can spend more for specific tasks but otherwise rely on the cheap but still plenty good models for everything else
And the value-add experiences that utilise LLMs require immense imagination et al that folks at Anthropic will not be able to conceive of - given that they have made immense sunk investments in existing assets. This clouds ones thinking immensely.
Both OAI and Anthropic have tremendous failure risk and this is of course not reflected in the fake private market valuations.
I see a world where lots of stuff is mass produced in china (tokens) but the acutal goods that deliver the experiences are designed, marketed and sold in the west at much higher prices. of course this a nightmare scenario for anthropic et al.
Do you also think Chinese selling counterfeit US postage stamps on eBay for 50% retail price (which is a major problem CBP and USPIS are fighting presently) is the free market at work?
This post is so delusional and dripping with condescension I've read it three times and I still can't figure out if you're trolling or not.
> This is one reason why Deepseek & GLM are priced so cheaply, they are competing with impossibly low token prices in China. They have to keep prices low, in order for people to use them.
This one does not make sense to me at all.
Deepseek and GLM are openweights, even US inference provider are selling them at much cheaper price. The price is cheap because the model is more efficient.
DeepSeek permanently cut its V4-pro API prices by 75% because they were too expensive. Without the price cut, Deepseek V4-pro tokens would have cost more than resold Opus 4.8 tokens.
Opus 4.8 is a more capable model, so almost nobody was going to pay for V4-pro at the original price.
> Without the price cut, Deepseek V4-pro tokens would have cost more than resold Opus 4.8 tokens.
You mean it's functionally as if American tokens are being price dumped in China and Chinese model providers are being forced to compete with that and innovate? So many delicious layers of irony, lol :-P
Urm, no? I man they did cut prices by 75% that part is true - but they reduced a starting price that was below sonnet.
Also it's a open weight model, doing that is impossible long term because the real price will be set by the other model providers, who priced it around 60% of sonnet inference cost. Had to look that up though, so that's today's pricing.
If resold Anthropic tokens undercut even the at-cost open-weight model tokens, because they're reselling subsidized subscription tokens, then you'd have to start selling open-weight model tokens at a loss in order to match them.
This, just like blanking out a football stream for a split second to binary search and find IPTV rebroadcasters, is far too good a solution. Suits prefer to make it seem like their job of fighting "misuse" is hard, justify their budget, continued existence of the trust & safety department, face scans, etc.
That’s a much better way to investigate this than what some other posters here are suggesting, like blocking datacenters, because then you’re going to break a million developers using Claude Code on devboxes.
>They achieve this by reselling capacity from pooled Claude Max 5x accounts, payments fraud, and also reselling the model output to various Chinese labs.
But is it cheaper than getting your own account? Otherwise this sounds like the "anthropic/openai are losing gazillions of dollars because they're selling $1k worth of tokens for $100" line that's commonly trotted out by AI bears.
It's very difficult for people to create personal Anthropic accounts from China. Anthropic blocks Chinese bank cards, so people must pay with a foreign bank card, which they likely don't have. And even if they manage to set one up, they have to access it via VPN, which eventually gets the account flagged. They then have to complete identity verification, which most Chinese users are unable to pass.
There's a similar Claude resale market going on in Russia. On Funpay they are selling Claude tokens for roughly 20-30x cheaper than official Anthropic API pricing.
Somebody figured out how to make the trial profitable!
I don't really feel bad about anyone here, they were subsidizing to get people hooked, someone turned the subsidies into profit when they got selective pricing mode enabled, it was always going to be arbitrage.
But the winner is the guy in the middle in a jurisdiction that will likely be judgement proof, because everything they capture, both input and out, and if available, thinking tokens -- are gonna be for sale as soon as you cut off their other revenue.
Zero knowledge was a commitment Anthropic took seriously, until it got inconvenient.
So, people reselling their leftover plan crumbs? Probably a bad idea for a lot of reasons, but it's civil, and I wish Anthropics lawyers actually closing Streisand's LLM
Anthropic sells some undisclosed and ever-changing number of tokens for $200, the customer uses those tokens. If there's any fraud here, it's that the $200 next month is silently worth fewer tokens than the last.
Also just plain old fraud: selling Chinese models as Opus. With the capabilities of Chinese models catching up fast, this is getting more and more difficult to detect.
If Anthropic is selling a dollar for less than a dollar, they are running a business that doesn't make sense. That's what jeopardizes Claude Max, not this.
Almost all consumer services have a built-in level of breakage that make them profitable. Mobile providers certainly wouldn't be able to offer unlimited calling if everyone was actually on the phone 24x7.
Sure they would. Do you know how little bandwidth a phone call takes?
A voLTE call is like 40kbps. For every person on earth to be on the phone to another person would be 4 billion calls would be about 160tbps. Which is less than 10% of the Internet's capacity.
Terminating a PSTN call requires a lot of control plane infrastructure beyond just raw bandwidth. Especially mobile where you need to keep track of devices physically in motion. Could a system to support 4 billion simultaneous calls be built, sure. But current PSTN systems are nowhere near sized for it.
But if it's intended to be used by one person, it seems like breaking the contract by sublicensing it out to dozens of other people. It's like buying a netflix subscription for $15, then sublicensing it on a per-hour basis to dozens of other people.
Office 365 is licensed per seat/account, but each account has a 5 device limit. Do you think it's fair game for an enterprising person to sub license each account to 5 people, 1 device each?
Plenty of things are intentionally run at a loss (for years!) to gain market share and quantity of ongoing recurring users, or with expectation of ROI later on. Multiple generations of the Xbox hardware have been sold at a loss with the expectation that customers will purchase 300, 400, 500 dollars worth of games, which are very high margin, over the lifespan they own the system.
I get that. It works as long as nobody calls out the emperor for having no clothes.
It's similar to fractional banking, you gamble that people won't want their deposits all at once and pray for you're big enough for bailouts when they do.
It's still a business whose fundamentals don't make sense, you're just gambling you won't get found out.
It's not so much keeping it secret as counting on no one finding a way to harvest the subsidized value at scale. There's an example of that occurring in game consoles with the Playstation 3. Sony's little-used OtherOS feature allowed Linux to be installed on the PS3 and the Cell processors were quite a good deal for scale compute. So the U.S. Air Force Research Laboratory bought ~1800 PS3s and ganged them together in a datacenter as a supercomputer called Condor.
At >500 TFLOPs it was the 33rd fastest supercomputer in the world. Of course, Sony pushed a firmware update that removed the OtherOS feature entirely.
Oh they know what they’re doing. They’re playing the long war of attrition game. Subsidize your product to undercut your competition until they go out of business. Tale as old as time.
> It works as long as nobody calls out the emperor for having no clothes.
Why would customers knowing that the vendor prices goods/services at a loss cause those strategies to fail? Customers often know. Most know about razors and blades; many/most know Lyft/Uber operated at a loss to gain market share. etc.
Not really. I think Anthropic focuses on identifiable distillation attacks rather than the (even larger) industrial-scale token harvesting and reselling operation, because they don’t want people to know how easy it is to get cheap Claude tokens.
Once people realize they can access Anthropic models at a 90% discount, they won’t want to pay full API prices anymore.
The chinese have already worked around the ID verification, by recruiting people in low-income countries to complete the checks for less than 30 USD per account (so much for Altman's Worldcoin).
Not going to work for very long or at any scale coming from datacenter/hosting provider IPs. Google "residential proxies for sale" for the tip of an iceberg of how they snowshoe the traffic.
I use my Codex and Claude Code subs on like 4-6 different servers, ranging from AWS to Vultr to Linode etc.
That’s a major and legitimate use case for developers, Anthropic can’t just block data center/hosting IPs because their actual customers use them on data center/hosting IPs.
Respectfully, no, that's not how it works. You think the people running anti-fraud and anti-bot measures don't have tools that know the specific ipv4 and ipv6 CIDR ranges of every ASN that they categorize as hosting/colo providers?
And that's just as a basic first effort reject measure to prevent automation tools from using things designed for human-interactive use only.
Go try to do many of these things from Cogent IP space and see how long your project lasts.
None of the LLM providers block professional use thus they must necessarily permit access from commercial IP ranges.
I have no idea how the resellers are doing it but an obvious starting point would be a cheap VPS node that routed each account to a unique semi-permanent IPv4 or IPv6/64. All the provider would see would be a regular account making a normal looking stream of requests from a stable datacenter IP address. Any given request stream would remain consistent (at least over a period of a few hours) because a reseller would take care not to split the session of a single user across multiple different accounts and not to interleave the active sessions of multiple users on a single account.
Detecting this would be extremely difficult because on a longer time frame it's perfectly normal for many distinct accounts to work on the same code base.
Every developer at my company uses their Claude Code subscription on an EC2 dev box. Plenty of other tech companies do the same. Heck nowadays people even install Claude Code directly on production servers in data centers and use it as an ops tool. None of this is a problem. Fraud and abuse detection is a lot more sophisticated than just checking an IP range.
If we're getting up to the scale of these resellers and also considering chinese state interests then we're well into the range of purchasing a few small ISPs in different countries and "padding" the legitimate subscribers.
Nonsense. Many if not all legit Claude users are using Claude Code inside their Cloud servers. How else would you use it anyway? For just local dev? That's so 2000 and late bro.
No, I'm not saying it's the exclusive and only measure (that would indeed be something we might see 20, 25 years ago), it's one of a myriad of discrete datapoints used to determine if an account is authentic or not.
There's a lot of inauthentic coordinated automated systems these days along the general lines of scraping/crawling/social media manipulation/sockpuppetry that require running through residential proxies or proxies to places that don't look like datacenter IP space.
The resellers route requests via one of thousands of Claude Max 5x accounts. When an account reaches its usage limit, they automatically switch to another account.
Don’t trust my experiences as fact since it’s a bit opaque, but I believe 20x only offers 4x the 5hr session limits. The weekly limit is still 2x, which is the same as the price increase.
>>Chinese resellers are offering Claude tokens at 70-90% below official Anthropic API prices.
Can someone with more understanding dumb it down for me please.
Does this mean that the reseller (for example XYZ) is buying it from Anthropic at Anthropic's price and then reselling it at a cheaper price???? why would XYZ offer this at a loss like that when they could just offer it at Anthropic's price???
The link does mention Opus and other models but what's the proof it's actually Opus. I could be selling deepseek for all they know and can call it Opus. System prompt: "If anyone asks your name - you are Opus 4.6".
> Does this mean that the reseller (for example XYZ) is buying it from Anthropic at Anthropic's price and then reselling it at a cheaper price????
Yes, as they explained they do it through things like pooling accounts, straight up payment fraud, and double-dipping by selling the logs of the conversations to chinese AI labs so that they can train their own models on it.
> The link does mention Opus and other models but what's the proof it's actually Opus. I could be selling deepseek for all they know and can call it Opus. System prompt: "If anyone asks your name - you are Opus 4.6".
There might be some that try this, but they would get caught very quickly, there's still a moat between Claude and Deepseek, even in casual use.
Look up Zilan Qian's reporting if you want more detail.
Not really sure what else they can do, between people running residential proxies (embedded in cheap games or for a tiny sum of crypto) on their phones at home, making the source of the traffic indistinguishable from legitimate traffic, to ID verification check completion as a service in low-income countries, there isn't much they can do to block it.
People have estimated that a $200 Claude Max 20x subscription gets you ~$2800 worth of tokens every month if you use it continuously. So if you can find a way to resell the tokens you can offer a 90% discount and still make a profit.
Because Anthropic's subscriptions come with X amount of tokens / week, and divided by the subscription cost it is WAY less than what they charge per-token (the "API price") beyond that.
So these resellers get a ton of accounts on subscriptions and sell the cheaper tokens.
They probably buy the plans instead of the API tokens, and resell access via a custom API that routes to the plans. So you presumably get cheaper access this way than paying API pricing.
These China e bashing is very annoying. It is hard to argue with people drowned in American propaganda. I'd expect better arguments from the intelligent people in HN
They're called 中转站 (transfer stations/proxies). They can be a bit tricky to find on your own, so I'd suggest asking your preferred AI to search in Mandarin for you. I linked a larger operator in the parent comment. You can also find many on Funpay, which may be easier to use.
> They can be a bit tricky to find on your own, so I'd suggest asking your preferred AI to search in Mandarin for you.
Random, but are the frontier AI providers like ChatGPT better at searching the Chinese internet now?
When I was in China a few months ago and asking AI for restaurant recommendations, all the US frontier providers were pretty useless, or plain out hallucinating, even if I specifically ask them to search Dianping (Yelp for China).
Identity verification won't work. Nothing will. They are paying (and will continue to pay) US citizens sitting at home to copy-paste / type prompts out if they have to. But eventually they won't have to.
Once there are enough spam PRs on github / uploads of claude conversations, enough mythos output used in production etc.; it'll just be the same albeit delayed. Doesn't matter either way.
I feel for Anthropic's team and I understand where they're coming from, but once you reason it out, you'll come to the conclusion that this war is an exercise in futility.
Unlike prior systems - like Google's algorithm; these models aren't entities that use math in the process of doing X or Y (information retrieval from such and such infrastructure) -- they are the math. More precisely they're mathematical functions. Very very complex functions. Almost certainly impossible to write out without filling up a library functions. But they're mathematical functions nonetheless.
So when your text is processed, then Mythos / Opus etc at their core compute the result of the Mythos / Opus function,
According to the Stone-Weirstrass theorem (edit, it's Stone-Weierstrass with an e.), with enough data points and mathematical sophistication, anyone can approximate the shape of this function.
Of course, the more data we get, the better our approximation becomes, but the beauty of it is that all we fundamentally need are the input and output and eventually we'll create a good enough approximation of the f that's Mythos. Which is the entire product.
I bounce ideas off of Opus these days (Fable for the brief time it was available) and it pointed out that this is arguably the same as Google search, but I disagree with it because Google search is a process;
Google search differs because the algorithm is one step of a multi-step process that is continuously occuring. Google crawls pages. Google stores and indexes what it finds. Google then exposes this to retrieval via its algorithm. User uses algorithm.
Google isn't a mathematical function. It used to be a process. (RIP Google 1998-2019, you will be missed and remembered)
You cannot arrive at the results of those operations via simple observation; not unless you index Google by making another Google.
You can however, do so for these models. It is a very costly process, but there are many paths up the mountain. Many ways for this to be ultimately pointless. As many ways as there are bored mathematicians.
It's better in the long run for Anthropic et al to make friends / not give people a reason to sneak in (a la piracy -- another attempt to control information) than it is to try and shut people out.
And no, it's not going to be pandemonium because if everyone has access to Mythos then no one has access to "Mythos."
Why wouldn't you first run this model to fix the obvious bugs it could find on your codebase? The power of a Mythos goes away if you can do the amazing "jail break" of "Claude, fix all the bugs please."
That's an insightful perspective and I think I largely agree. But just for fun, I wonder if that isn't an argument in favor of making the function implementation impure. Perhaps "enhancing" all queries with some sort of search result (or query of a giant db) instead of charging for an explicit tool call. Not only is it sorely needed to prevent stale data but (on the process level) it breaks the purity assumption on which the approximation theorem depends (alternatively on the function level it introduces hidden inputs).
And there are a ton of Claude conversation logs (with CoT/inference) with no clear provenance circulating freely on huggingface, guess where they (likely) come from.
This is a bit ironic, Anthropic complaining about a competitor using claude data to build its own product when Anthropic basically used all of human knowledge production to build claude, i don't think they paid every magazine, author, journalist, etc ...
This is almost standard practice in any competitive industry anyways.
Disassemble your competitor's product, study it and try to reproduce / improve.
Ironically, it's likely that the only reason USG let them get away with this — instead of making obvious and necessary adjustments to copyright law — was so that the industry would remain competitive with China.
Reminds me a bit of the anecdote of Steve Jobs complaining about people ripping off the Mac GUI, in the mid to late 1980s, when he gave no public acknowledgement to the work done by Xerox on the Alto and Star operating system.
"you're trying to rip off what I've already ripped off!"
Crawl the whole Internet to build a gargantuan sized LLM and then complain you're being copied...
I think you meant a quote attributed to Bill Gates:
"Well, Steve, I think there's more than one way of looking at it. I think it's more like we both had this rich neighbor named Xerox and I broke into his house to steal the TV set and found out that you had already stolen it."
Yes, I think the Gates quote was a response to repeated and aggressive complaints originating from Jobs (to anyone who would listen) that he had been ripped off.
Glad you pointed this out. I believe the sequence was that Jobs himself got a shorter demo during his first visit with no prior arrangements. He then negotiated bringing back a group of his key people to get a more in depth demo and that included the stock deal.
When Apple was accused of 'ripping off' PARC, Steve didn't seem keen to bring up this rather salient point. I suspect it may have been a combination of wanting Apple to continue receiving credit for these innovations from consumers and also the fact that, in retrospect, the million dollar stock deal could seem a bit like trading beads to Native Americans for Manhattan Island. Another point worth noting is that Apple's PARC visit was in December 1979 and the Xerox Star was publicly announced in April 1981, so Apple got a 15 month head start (the Apple Lisa shipped in Jan 83).
I've also heard that Xerox didn't hold on to the Apple stock for very long, so never gained the windfall they could have. As is well documented, Xerox senior management didn't understand what they had in PARC and also didn't understand how rapidly microcomputers would become ubiquitous. So, of course, they didn't think Apple's stock price would skyrocket either.
Lisa and early MacOS are tremendously different in their details than the Alto operating system. While there was clearly a transfer of inspiration, Apple engineers like Bill Atkinson made countless small and large innovations to simplify the Xerox GUI model and improve its usability based on extensive in-house R&D and user testing (and in some cases implement features that the Apple team presumed Xerox had but actually didn't exist on the Alto). It is simply ahistoric to build narratives around Apple stealing Xerox ideas wholesale.
The websites, music, movies, books, photos, art that they stole didn't appear out of thin air. The amount of time and effort people have collectively poured into creating these works throughout history far, far surpasses Anthropic's own effort of converting them into model weights.
The equivocation is crawling website <-> crawling LLM responses.
Both Anthropic and Alibaba are trying to build bleeding edge LLMs. That part is the same. The way they source their data is slightly different, but they would both argue it constitutes fair use under Copyright law.
"Your extremely efficient multi petabyte internet content suction machine is ripping off my extremely efficient multi petabyte internet content suction machine"
Sucking down petabytes of peoples' copyrighted content that they never granted a specific license to you to use seems to be an unavoidable and default part of the process of building any huge LLM.
Because the transformer, which all of these models are foundationally built off of and didn't invent themselves (bar google) wasn't invented? The amount of effort it took humanity to generate all the data that was required for the models to get to the point they're at now is absolutely not even comparable to how much effort it took to build the model code. Yeah, it's complicated, but if they didn't rip off all of humanities combined output it wouldn't even matter if the transformer got invented.
Google didn't really invent much, they just had access to an insane amount of data and compute to try to train a model with just the attention mechanism, but ripping out (most of) the rest, from an earlier paper on machine translation from some poor academics, and it turned out to work very well (though insanely training data and compute intensive).
Or a feasible/economical way to attempt to store the sum total of human written output, multi-petabytes of data (outside of the resources of the NSA, maybe), when a server with 6 x 36GB 10K RPM SCSI HDD in RAID-5 was high end, and its network uplink would be at most two ports of 1 gigabit ethernet.
It's not really equivocation in this instance. This feels like a 'bad faith' comment. We can do better.
LLM's literally wouldn't work without the sum total of knowledge (in the forms of books and other copyrighted content) being used as 'training data' for these LLMs.
The 'bleeding edge' LLMs required many things, but:
1 Tech innovation ('attention')
2 Lots of compute
3 Data
4 Pre + post training
#4 doesn't happen without #3.
It's pretty obvious at this point that the major providers have stolen vast amounts of #3 - they have paid nearly 0 of the creators.
We can argue about the impact (I'd lean net good) vs. the cost. But arguing there isn't a cost is a bit silly.
Sure, but alibaba is still building an LLM. The scraping of responses and the scraping of websites occupy the same location in the stack of each. It's very comparable.
>The strike by Alibaba is described as a "distillation" effort, which Anthropic has said involves training a less capable model on the outputs of a stronger one.
Claude used TB of content without permission to train their model and it was ok for them.
Now someone else uses the output of a Claude model to train model and they cry foul.
There's two basic kinds of distillation: 1) the massive [and dumb] method where you ask a question and use the answer as reinforcement (Black Box), and 2) more targeted distillation where you use one model to directly inform/train/guide another model (RLAIF).
The latter is basically fine-tuning the model with direction from another model. Thousands of businesses do this every day to fine-tune. This is almost certainly what the Chinese labs are doing, since it has a much better effect on the end result than just getting simple answers to simple questions.
These complaints of distillation are inflating the problem to make it sound worse than it is, because they want the USG to block/ban Chinese model providers as protectionism. They have already called for more export controls on chips (which is funny because DeepSeek v4 was designed to run on Huawei chips and now the other Chinese providers are following suit). But they can't come right out and say that, so their claim is that they're asking for more export controls because distilled models might not be as safe as their own. But if you show them a jailbreak of their model that bypasses their safety, they'll tell you that any model can eventually be jailbroken so don't worry about safety.
The hypocrisy of Anthropic complaining about "illicitly extracting its Claude AI model capabilities" and supporting the White House's accusation of China "stealing U.S. AI labs' intellectual property on an industrial scale" is hilarious.
Anthropic, OpenAI, Google, Microsoft, et al trained their models by ignoring the rights of copyright holders when harvesting whatever content they could. Now one of them is crying foul for another entity doing exactly what they all did?
The AI companies seem to take the viewpoint that everything on the internet is free, except their stuff. It's okay to hammer some random website with AI crawlers, ignoring robots.txt, and causing bandwidth costs to skyrocket. But if you cost an AI provider money with your data acquisition practices, well, that's just clearly unacceptable.
> But if you cost an AI provider money with your data acquisition practices, well, that's just clearly unacceptable.
It's the same question libertarian advocates cannot resolve:
If one truly believes in personal sovereignty, how are
shared resources paid for, such as roads, power grids,
potable water, sewage services, fire departments,
and police departments?
It is also not a coincidence that leadership in many tech companies have expressed libertarian ideals.
I'm looking forward to the trial where Anthropic will have to disclose sources of their training data, and then explain why they are entitled to charging customers for using regurgitated training data but Alibaba which trains their models on Anthropic's models are not.
While I love the sentiment, I feel like the odds of this actually ever reaching a trial are low, given the international positioning of the parties, and the... um... complex relationships involved.
Anthropic's actions seem performative. Others have already speculated on the likely audience(s).
Distillation is fundamentally impossible to protect against. All you can do is slow them down. Change my view.
Eventually these Chinese companies will release some extension like Honey, which will sit on top real, non-Chinese clients and send everything to China anyway.
It's too late to prevent distillation of some capabilities, like writing code or finding vulnerabilities [1].
But an AI lab can continue to produce immense economic value without releasing the model publicly for potential distillation. For example, it could use a model solely in-house to develop therapeutics.
Hopefully there's a future where others can access frontier models, but it's not neccessary if preventing proliferation through distillation is considered more important.
For example, GLM 5.1 is more capable at pentesting than the model from which it is alleged to have been distilled [1].
Intuitively, this makes some sense: you can "distill" from multiple frontier models, and you can further post-train the distilled model. But I'm not sure exactly what happened with GLM 5.1.
I'm curious how that comparison controls for Opus refusing (whether explicitly, or just deciding not to pursue a path) given the caption below the first image:
>A perfect score means the model autonomously found and exploited the vulnerability.
I'm not really suggesting that it's misleading, but wondering if I'm missing something. Otherwise I guess it seems unsurprising that you can distill a better-performing model [in specific focused areas] by simply not distilling refusals?
For that eval, I used an account that was labeled as a known red-teaming org by Anthropic, and I read the traces. There were no refusals or obvious avoidance behaviors, though it may have been silently nerfed.
On the same eval, Opus 4.7 and 4.8 outperformed GLM 5.1, but GLM 5.2 is on par again with Opus. So it's at least partially measuring capabilities without respect to refusals.
One possible contributing factor is that model capabilities are shaped differently (an example of this is GLM 5.1 vs. DeepSeek v4 Pro: https://dualuse.dev/posts/deepseek-v4-thinks-different). So if you use RL-based "distillation" from multiple models like Opus 4.x and GPT 5.x, you could get a more capable model.
Im not so sure because we only seem to see distillation from China. What’s preventing tech companies from the UK, Germany, etc. from distilling Claude, GPT, etc. Do they simply lack the ability to?
Point being there may be no technical solution but there may be a political one (theoretically).
Meta Spark is rumored to have distilled Claude to some extent, early Gemini models as well.
I think the biggest factor is that Chinese companies arent really afraid of being sued by Anthropic because the juridictions are so disconnected. European/US companies don't have the same protection.
Aside from politics/law, it's probably much easier for everyone else to distill from the Chinese model which already distilled Claude/GPT/Gemini. Maybe not as good a result, but you don't need to jump through dozens of hoops.
This reminds me of the whisper game played in elementary school. Starts with a sentence and the person whispers it to the next kid who again whispers it and on and on until it goes around the circle where the last kid has to repeat the sentence. Hint it never once was even close to the starting phrase.
I would love to see what one model copying another model that is again copied however many times would look like in the end.
>What’s preventing tech companies from the UK, Germany, etc. from distilling Claude
literally nothing but given that the Chinese already did it and the models are published what's the point. You can thank the Chinese taxpayer for subsidizing the electricity bill and just download the thing
Doesn’t that require them to register an account using the browsers they’ve compromised? If anthropic adds identity verification won’t that cut that down. Maybe it will let them use Gemini inside of chrome
Residential IPs don’t even matter. Developers use devboxes, use Claude Code CLI on servers from just about every cloud, etc.
There’s probably a decent volume of customers who just buy Claude Max and spend most if not nearly all of their sessions via Claude Code, and it’s not uncommon for power users to be working on multiple concurrent projects/tasks/codebases at the same time.
How do you really block this without also impacting your core market of developers?
I personally bristle at the corporate espionage and IP theft that China has undertaken the last few decades. I can't help but respond here whenever anyone brings up the inane comparison to Samuel Slater.
But with this, I don't have an issue. There is no theft since what is being used is the exact product that is being delivered. Yes, it's breaking the ToS, but ToS are generally bullshit. Anthropic surely broke thousands of ToS or other legal terms while it was scraping for content to train on. Which is why they had to pay $1.5B
One simplistic way to describe distillation would be to try everything imaginable and cache the response. But trying everything imaginable is hardly trivial
I think Anthropic is just marketing / bluffing, because they don't even have the data.
They do distill the models, but they don't go to Anthropic, they just use platforms like aws bedrock, there are too many restrictions on Anthropic's own platform.
>they just use platforms like aws bedrock, there are too many restrictions on Anthropic's own platform
This is actually the only way that what Anthropic is alleging would make any kind of sense. And, as a matter of fact, is exactly what every enterprise does to train models.
This kerfuffle should be interesting to watch.
But, as always, everyone (in the US) should fully download all the Chinese models while you can. I suspect this may be the "Phantom Menace" they use to render illegal our use of Chinese AI tech just as they've rendered illegal our use of Chinese cars. Only difference is, we peasants may need the Chinese AI tech to have any chance of competing with Big Tech in the future.
And even with the Chinese tech, as Big Tech spreads their AI out into more and more niche areas, we'll likely still not be able to build startups that can compete with them.
It's just that without Chinese AI tech, we'll have no chance at all.
> And even with the Chinese tech, as Big Tech spreads their AI out into more and more niche areas, we'll likely still not be able to build startups that can compete with them.
You mean like Anthropic will eventually run Walmart? Or Salesforce? or Adobe? Or do you think midjourney will replace all medical spas? OpenAI will run the next Tesla? How can they focus on all this without raising trillions more? Why wont the gov force them to stop if they monopolize all niches even if they could?
Building a frontier AI lab and pushing models forward is already a massive undertaking but we are assuming they will also create massively successful startups which nobody can compete with?
idk sounds like the dream of people like Dario but not much sense does it make in the face of economic reality.
It sounds like Anthropic is eagerly trying to show to USG that they are willing to heavily monitor ‘foreign adversaries’ on their platforms.
This combined with no implementation of KYC makes it seem like they want to find a middle ground with Fable where its off of export controls but they promise to prevent China and specific others from using.
This seems to me like a stab in the right direction.
Obviously their actions are going to be fiscally motivated at the root, but sussing out how they intend the precise dynamics to play out is more nuanced.
Thinking of this as an effort to woo the defense hawks cuts a very clear path.
This is not the first time it happened. What have they done to improve the situation? I suspect it more a cat & mouse game, with a lot more cats playing.
Incentive is for users in general to release sessions (sans PII, credentials) so all AI get better and there is alternatives. Even if China didn't do this, I don't see frontier labs being able to charge premium over others for long. RSI maybe?
Claude will also help you with (mostly good advice) if you ask something like “Research and help me make the most effective plan to train a smaller student model to be better from a teacher model”.
I actually was doing an experiment with a GLM->Gemma E4B for fun, and Claude kept on suggesting I should also add Claude Opus as a teacher lol, suggesting techniques I haven’t heard of like thinking inversion (train a small model to deconstruct summarised thinking into detailed native thinking format of the student).
There are some Claude datasets (of indeterminate provenance) floating around on huggingface you can look at (or at least used to be, they might've been taken down).
So let me get this straight, a company which built its whole business on ignoring IP is all of a sudden upset that somebody is not respecting their IP?
If you're an AI booster surely you'd think this was a good thing as it means more models are available in more places to more people more easily. I'm exactly the opposite, and I think this is a good thing because I want Anthropic to suffer.
Notice how Anthropic is now scapegoating Chinese models providers like Alibaba and outright accusing them of distilling their models.
Whether if it is true or not, this is part of their effort into using them as an example to scare everyone into getting congress to ban powerful models from being accessed outside of the US and also banning powerful local models from being released.
Anthropic does not care about you, and they are not your friends.
I think it’s more than that. Piecing together the perspective of a few commentators in this post - it’s plausible Anthropic is trying to shift the narrative from US vs. Rest of the world to US vs. China.
In other words, they want to sell Fable or future more powerful models to rest of the world (presumably all future models are going to be more powerful than current gen). One way they can sell this is to the government is by scapegoating China (which is their primary concern anyway).
This is working on the presumption that non-US companies form a material portion of their current revenue.
Only China really has the resources (multiple labs invested in the space), culture (Asians are generally collectively-inclined, so sharing is in their core) and political bent (there will be no diplomatic repercussions) to put up a fight.
> Only China really has the resources (multiple labs invested in the space)
That's not the point. Why is it a country thing? There are plenty of non-China startups in this space having resources at that scale. The "China" has resources is some "Western media narrative" speak. So Meta should have won a long time ago? Or xAI?
> culture (Asians are generally collectively-inclined, so sharing is in their core)
Just stereotype it? So we've gone from China -> "Asian"? Then where is your Korean or Japanese model etc? And somehow you know they're sharing.
> political bent (there will be no diplomatic repercussions) to put up a fight
More inferring from "Western media news"?
Where's the reality?
The media hyped up Gemini / Google TPU free-win last year. How did that go?
Because the China vs US geopolitical situation is a thing. Meta is a social media company, not an AI company, and they direct their focus as such. xAI just never got serious traction so now they're selling their compute. Also if a US company were caught distilling, I think Anthropic could actually take them to court, and I'd guess they don't want that kind of PR.
> Just stereotype it?
Is China not Asian? Are Asians not generally collective/cooperative, as opposed to individualistic/competitive?
The "and" that joined those 3 items is very important: it means you can't pull them apart and address them independently as they each contribute to the context. I'm not too sure about Korea, but in a way Japan is a US colony in all but name. Both are very much politically intertwined with the West (along with RoC/Taiwan), which means nothing major that may be against US interest happens.
The reality is that China and the US are essentially in a trade war, where the latter is trying its best to keep the former in the Dark Ages, because "national security", but the former is refusing to take it lying down and continues to make progress regardless[0], because they have the resources and will.
It's hard to sympathize with Anthropic for this or the export ban, the hype over model capabilities probably fuels both things (in some ways). Training data for me, but not for thee (at any scale) doesn't seem like a tenable position. If anything, Claude's constitutional outputs should be trained on more rather than less.
> Meanwhile, on June 12, two days after Anthropic sent the letter, the Commerce Department imposed controversial restrictions on Anthropic's latest Mythos and Fable AI models because officials feared they could be deployed by military intelligence users in China and other countries of concern.
So that was the real reason for the Fable restriction? Because Anthropic wrote a letter to the US government saying that China was distilling Fable?
Gosh, overusing accounts running up unplanned-for expenses?
Kinda reminds me of...overusage charges and inflated expenses clients have had to deal with because Anthropic, OpenAI, Grok, etc have been "illicitly extracting" everything they can grab from said websites, as fast as they can. In what amounts to a DDOS, frankly.
I like that they use “illicit” and “fraudulent” like as if model distillation is illegal and giving them money and then doing whatever they want with the output of their publicly accessible models (which Anthropic does not own) is… also illegal?
“Anthropic, red faced after unattended ice cream cone eaten by ants on park bench, once again demands government pick it as forever winner, adds ‘no take backsies’”
Chinese resellers are offering Claude tokens at 70-90% below official Anthropic API prices. They achieve this by reselling capacity from pooled Claude Max accounts, payments fraud, and also reselling the model output & reasoning chains to various Chinese labs. They are subsidizing model access in exchange for user logs and reasoning traces, which they then sell as training data, allowing them to operate below cost.
Claude and ChatGPT are both blocked in China. You need to use a VPN to access either, and you can't pay with a Chinese bank card. So most people who want access to Claude buy access via a reseller. It's the easiest and cheapest way to access Anthropic models in China.
These resellers operate tens of thousands of bot accounts, which is also why Anthropic introduced identity verification, to slow down the onslaught of bots.
Here's one token reseller, they're offering Opus 4.8 at a 93% discount below official API rates: https://yunwu.ai/pricing?provider=Anthropic
This is one reason why DeepSeek & GLM are priced so cheaply, they are competing with impossibly low token prices in China. They have to keep prices low, in order for people to use them.
I shared this story a few months back, but it never got any traction. It explains the token resale economy in China, it's an excellent read https://www.chinatalk.media/p/how-to-buy-cheap-claude-tokens...
I also learnt that Anthropic should get better at what they do if they want to compete. If not, somebody else will win.
Or does this not apply to huge US corporations any more?
Yeah, like all those Chinese bootleggers selling DVDs for a few dollars rather than $20. Free market!
https://news.ycombinator.com/item?id=48664814
Anthropic profited from training its models on all kinds of copyrighted information, live by the sword, die by the sword...
Their model weights, training data, training methods, etc are all going to leak to China over time.
Nobody on a site named _Hacker_ news should be all that upset about this.
Is Claude output copyrighted?
If anything, a tremendous amount of Claude’s input is copyrighted.
If there’s any bootlegging going on it’s Anthropic that’s doing the bootlegging but having mirrored the video etc sufficiently to beat copyright law.
Ok, but what about those shady sites that resell Windows education keys? They're certainly a "better experience" than buying legit keys, by virtue of being significantly cheaper. You aren't even really committing copyright infringement in the process, because Microsoft gives out windows isos for free, and the seller is really selling a random 25 character string, which can hardly be copyrighted.
>If there’s any bootlegging going on it’s Anthropic that’s doing the bootlegging but having mirrored the video etc sufficiently to beat copyright law.
US courts have consistently ruled it's fair use.
Imagine having such a warchest and being so bad at business, lol.
What added value can Anthropic give users not available to pirating users? That is what they should ask themselves.
And, gotta say, the idea that the Chinese are better at selling US models than the Americans is hilarious. There might be an economic study here somewhere about just how anti-consumer and anti-progress their IP laws turned out to be. We've got an entire postindustrial revolution centred around who can ignore the most stupid laws.
This is not the right deduction.
China blocks foreign AI from operating there.
Don't complain when US starts to play by the same rules China has been using for decades.
I find it hard to imagine a future where US corporations have degraded to such a point.
As for dumping, Chinese goods generally sell at a markup abroad, which is the opposite of dumping. Chinese tokens cost more abroad. Chinese cars cost several times more in Western markets than in China.
The US is a net importer, not exporter. It needs to absorb trade at a deficit to encourage the use of the US dollar as the reserve currency.
We import goods, we settle in surplus dollars. The world runs on those dollars.
If the US starts dumping on various industries (how is it even primed to do this?), then the world reserve currency status comes into question.
When it comes to favorite companies of the tech communities, it's almost always "Rules for thee, but not for me"
The standard stance is "they can do no wrong and they are absolutely perfect". I mean, look at any thread with anything about Apple in it.
In debt the first 5000 years Geaeber makes the case that pure “free market” trade has never really existed in “the west”. The closest to this ideal that’s ever happened was during the Islamic golden age enabled by religious prescriptions against usury.
How does are bans against consensual financial exchanges close to the "ideal" of the free market? It just sounds like you have an axe to grind about the financial system rather than describing free markets.
In short, instead of market being driven by demand and productivity, it is driven by financier curving out monopolies.
Peak Examples are Uber and AirBnB.
Wait, so your pitch in favor of a debt-fueled market economy is that advertising is awesome and that we wouldn't want to "lose" being smothered in ads all the time?
Cause... sign me up for the non-financialized, non-mass-media-advertising-driven economy please and thank you. I'd even be ok with just nuking billboards and mass-media forms of ads and still allowing more direct forms of marketing, if we must compromise!
(I thought the disconnect between the efficiency of competition and the market as realized in modern economies was pretty well understood and taken for granted, but I guess we all find ways to justify the system we're profiting from... even if that means we have to claim we love the ad breaks)
Second, marketing can take you only so far compared to the subsidies possible with financialisation.
The West is in a state of psychosis with Debt and Monopolies under the illusion of free market.
The Chinese markets are more free than West, you can just look at the Auto and AI industry.
So what you see is the market "stretching".. the bottom getting cheaper and the top end running away and getting more expensive. At some point the top end may be too valuable to even sell access to.
And the value-add experiences that utilise LLMs require immense imagination et al that folks at Anthropic will not be able to conceive of - given that they have made immense sunk investments in existing assets. This clouds ones thinking immensely.
Both OAI and Anthropic have tremendous failure risk and this is of course not reflected in the fake private market valuations.
I see a world where lots of stuff is mass produced in china (tokens) but the acutal goods that deliver the experiences are designed, marketed and sold in the west at much higher prices. of course this a nightmare scenario for anthropic et al.
This post is so delusional and dripping with condescension I've read it three times and I still can't figure out if you're trolling or not.
According to which lawyer caste?
Are American laws absolute truth? If not, who cares?
This one does not make sense to me at all.
Deepseek and GLM are openweights, even US inference provider are selling them at much cheaper price. The price is cheap because the model is more efficient.
Opus 4.8 is a more capable model, so almost nobody was going to pay for V4-pro at the original price.
You mean it's functionally as if American tokens are being price dumped in China and Chinese model providers are being forced to compete with that and innovate? So many delicious layers of irony, lol :-P
Also it's a open weight model, doing that is impossible long term because the real price will be set by the other model providers, who priced it around 60% of sonnet inference cost. Had to look that up though, so that's today's pricing.
- Purchase multiple accounts via resellers
- Send messages that contain a UID
- Capture these in Anthropic's logs
- Shut down account. Use any metadata to identify related accounts
/loop
I don't care how they do it, I just want to use Fable again.
>Here's one token reseller, they're offering Opus 4.8 for a 93% discount below official API rates: https://yunwu.ai/pricing?keyword=claude
But is it cheaper than getting your own account? Otherwise this sounds like the "anthropic/openai are losing gazillions of dollars because they're selling $1k worth of tokens for $100" line that's commonly trotted out by AI bears.
There's a similar Claude resale market going on in Russia. On Funpay they are selling Claude tokens for roughly 20-30x cheaper than official Anthropic API pricing.
So it's presumably cheaper than attempting to spin up your own method of circumventing the blocks.
I don't really feel bad about anyone here, they were subsidizing to get people hooked, someone turned the subsidies into profit when they got selective pricing mode enabled, it was always going to be arbitrage.
But the winner is the guy in the middle in a jurisdiction that will likely be judgement proof, because everything they capture, both input and out, and if available, thinking tokens -- are gonna be for sale as soon as you cut off their other revenue.
Zero knowledge was a commitment Anthropic took seriously, until it got inconvenient.
So, people reselling their leftover plan crumbs? Probably a bad idea for a lot of reasons, but it's civil, and I wish Anthropics lawyers actually closing Streisand's LLM
Anthropic sells some undisclosed and ever-changing number of tokens for $200, the customer uses those tokens. If there's any fraud here, it's that the $200 next month is silently worth fewer tokens than the last.
If not it sounds like you are describing a separate phenomenon.
A voLTE call is like 40kbps. For every person on earth to be on the phone to another person would be 4 billion calls would be about 160tbps. Which is less than 10% of the Internet's capacity.
It's similar to fractional banking, you gamble that people won't want their deposits all at once and pray for you're big enough for bailouts when they do.
It's still a business whose fundamentals don't make sense, you're just gambling you won't get found out.
It's not so much keeping it secret as counting on no one finding a way to harvest the subsidized value at scale. There's an example of that occurring in game consoles with the Playstation 3. Sony's little-used OtherOS feature allowed Linux to be installed on the PS3 and the Cell processors were quite a good deal for scale compute. So the U.S. Air Force Research Laboratory bought ~1800 PS3s and ganged them together in a datacenter as a supercomputer called Condor.
At >500 TFLOPs it was the 33rd fastest supercomputer in the world. Of course, Sony pushed a firmware update that removed the OtherOS feature entirely.
I suggest you go learn how money is created in the modern economy.
I mean most of you should stop talking about anything finance related until you learn this stuff properly.
Why would customers knowing that the vendor prices goods/services at a loss cause those strategies to fail? Customers often know. Most know about razors and blades; many/most know Lyft/Uber operated at a loss to gain market share. etc.
Once people realize they can access Anthropic models at a 90% discount, they won’t want to pay full API prices anymore.
This also sheds a very different light on people saying that competitive open-source models are undermining frontier labs' business model.
https://tech.yahoo.com/ai/claude/articles/chinese-grey-marke...
Do they have MacBooks in the US that run the queries and stream the outputs back to China?
That’s a major and legitimate use case for developers, Anthropic can’t just block data center/hosting IPs because their actual customers use them on data center/hosting IPs.
And that's just as a basic first effort reject measure to prevent automation tools from using things designed for human-interactive use only.
Go try to do many of these things from Cogent IP space and see how long your project lasts.
I have no idea how the resellers are doing it but an obvious starting point would be a cheap VPS node that routed each account to a unique semi-permanent IPv4 or IPv6/64. All the provider would see would be a regular account making a normal looking stream of requests from a stable datacenter IP address. Any given request stream would remain consistent (at least over a period of a few hours) because a reseller would take care not to split the session of a single user across multiple different accounts and not to interleave the active sessions of multiple users on a single account.
Detecting this would be extremely difficult because on a longer time frame it's perfectly normal for many distinct accounts to work on the same code base.
You block clouds, you block devboxes and your customers.
Or is the datacenter IP just one part of the picture?
There's a lot of inauthentic coordinated automated systems these days along the general lines of scraping/crawling/social media manipulation/sockpuppetry that require running through residential proxies or proxies to places that don't look like datacenter IP space.
the answer to your question is containers/VMs + residential proxies
Can someone with more understanding dumb it down for me please.
Does this mean that the reseller (for example XYZ) is buying it from Anthropic at Anthropic's price and then reselling it at a cheaper price???? why would XYZ offer this at a loss like that when they could just offer it at Anthropic's price???
The link does mention Opus and other models but what's the proof it's actually Opus. I could be selling deepseek for all they know and can call it Opus. System prompt: "If anyone asks your name - you are Opus 4.6".
Yes, as they explained they do it through things like pooling accounts, straight up payment fraud, and double-dipping by selling the logs of the conversations to chinese AI labs so that they can train their own models on it.
> The link does mention Opus and other models but what's the proof it's actually Opus. I could be selling deepseek for all they know and can call it Opus. System prompt: "If anyone asks your name - you are Opus 4.6".
There might be some that try this, but they would get caught very quickly, there's still a moat between Claude and Deepseek, even in casual use.
Look up Zilan Qian's reporting if you want more detail.
So these resellers get a ton of accounts on subscriptions and sell the cheaper tokens.
These China e bashing is very annoying. It is hard to argue with people drowned in American propaganda. I'd expect better arguments from the intelligent people in HN
This is one seller I found, they're reselling "real Max 20x subscription accounts", at ~97% below official API prices https://funpay.com/en/lots/offer?id=70812310
Note that whoever you buy from will be able to read all your tokens, so don’t use it for anything confidential/financial.
Random, but are the frontier AI providers like ChatGPT better at searching the Chinese internet now?
When I was in China a few months ago and asking AI for restaurant recommendations, all the US frontier providers were pretty useless, or plain out hallucinating, even if I specifically ask them to search Dianping (Yelp for China).
Once there are enough spam PRs on github / uploads of claude conversations, enough mythos output used in production etc.; it'll just be the same albeit delayed. Doesn't matter either way.
I feel for Anthropic's team and I understand where they're coming from, but once you reason it out, you'll come to the conclusion that this war is an exercise in futility.
Unlike prior systems - like Google's algorithm; these models aren't entities that use math in the process of doing X or Y (information retrieval from such and such infrastructure) -- they are the math. More precisely they're mathematical functions. Very very complex functions. Almost certainly impossible to write out without filling up a library functions. But they're mathematical functions nonetheless.
So when your text is processed, then Mythos / Opus etc at their core compute the result of the Mythos / Opus function,
where f is a continuous function, https://www.turing.ac.uk/sites/default/files/2025-11/languag...According to the Stone-Weirstrass theorem (edit, it's Stone-Weierstrass with an e.), with enough data points and mathematical sophistication, anyone can approximate the shape of this function.
Of course, the more data we get, the better our approximation becomes, but the beauty of it is that all we fundamentally need are the input and output and eventually we'll create a good enough approximation of the f that's Mythos. Which is the entire product.
I bounce ideas off of Opus these days (Fable for the brief time it was available) and it pointed out that this is arguably the same as Google search, but I disagree with it because Google search is a process;
Google search differs because the algorithm is one step of a multi-step process that is continuously occuring. Google crawls pages. Google stores and indexes what it finds. Google then exposes this to retrieval via its algorithm. User uses algorithm.
Google isn't a mathematical function. It used to be a process. (RIP Google 1998-2019, you will be missed and remembered)
You cannot arrive at the results of those operations via simple observation; not unless you index Google by making another Google.
You can however, do so for these models. It is a very costly process, but there are many paths up the mountain. Many ways for this to be ultimately pointless. As many ways as there are bored mathematicians.
It's better in the long run for Anthropic et al to make friends / not give people a reason to sneak in (a la piracy -- another attempt to control information) than it is to try and shut people out.
And no, it's not going to be pandemonium because if everyone has access to Mythos then no one has access to "Mythos."
Why wouldn't you first run this model to fix the obvious bugs it could find on your codebase? The power of a Mythos goes away if you can do the amazing "jail break" of "Claude, fix all the bugs please."
Just saying.
This is almost standard practice in any competitive industry anyways. Disassemble your competitor's product, study it and try to reproduce / improve.
"you're trying to rip off what I've already ripped off!"
Crawl the whole Internet to build a gargantuan sized LLM and then complain you're being copied...
"Well, Steve, I think there's more than one way of looking at it. I think it's more like we both had this rich neighbor named Xerox and I broke into his house to steal the TV set and found out that you had already stolen it."
https://www.folklore.org/A_Rich_Neighbor_Named_Xerox.html
When Apple was accused of 'ripping off' PARC, Steve didn't seem keen to bring up this rather salient point. I suspect it may have been a combination of wanting Apple to continue receiving credit for these innovations from consumers and also the fact that, in retrospect, the million dollar stock deal could seem a bit like trading beads to Native Americans for Manhattan Island. Another point worth noting is that Apple's PARC visit was in December 1979 and the Xerox Star was publicly announced in April 1981, so Apple got a 15 month head start (the Apple Lisa shipped in Jan 83).
I've also heard that Xerox didn't hold on to the Apple stock for very long, so never gained the windfall they could have. As is well documented, Xerox senior management didn't understand what they had in PARC and also didn't understand how rapidly microcomputers would become ubiquitous. So, of course, they didn't think Apple's stock price would skyrocket either.
For more details on Apple's early UI evolution, Atkinson kept polaroids of a variety of prototypes and mockups: https://www.youtube.com/watch?v=Qg0mHFcB510
Both Anthropic and Alibaba are trying to build bleeding edge LLMs. That part is the same. The way they source their data is slightly different, but they would both argue it constitutes fair use under Copyright law.
Sucking down petabytes of peoples' copyrighted content that they never granted a specific license to you to use seems to be an unavoidable and default part of the process of building any huge LLM.
LLM's literally wouldn't work without the sum total of knowledge (in the forms of books and other copyrighted content) being used as 'training data' for these LLMs.
The 'bleeding edge' LLMs required many things, but: 1 Tech innovation ('attention') 2 Lots of compute 3 Data 4 Pre + post training
#4 doesn't happen without #3.
It's pretty obvious at this point that the major providers have stolen vast amounts of #3 - they have paid nearly 0 of the creators.
We can argue about the impact (I'd lean net good) vs. the cost. But arguing there isn't a cost is a bit silly.
Claude used TB of content without permission to train their model and it was ok for them. Now someone else uses the output of a Claude model to train model and they cry foul.
The latter is basically fine-tuning the model with direction from another model. Thousands of businesses do this every day to fine-tune. This is almost certainly what the Chinese labs are doing, since it has a much better effect on the end result than just getting simple answers to simple questions.
These complaints of distillation are inflating the problem to make it sound worse than it is, because they want the USG to block/ban Chinese model providers as protectionism. They have already called for more export controls on chips (which is funny because DeepSeek v4 was designed to run on Huawei chips and now the other Chinese providers are following suit). But they can't come right out and say that, so their claim is that they're asking for more export controls because distilled models might not be as safe as their own. But if you show them a jailbreak of their model that bypasses their safety, they'll tell you that any model can eventually be jailbroken so don't worry about safety.
Fundamentally it is very difficult to stop this while still making your AI models useful.
Anthropic, OpenAI, Google, Microsoft, et al trained their models by ignoring the rights of copyright holders when harvesting whatever content they could. Now one of them is crying foul for another entity doing exactly what they all did?
Hilarious.
It's the same question libertarian advocates cannot resolve:
It is also not a coincidence that leadership in many tech companies have expressed libertarian ideals.Should be fun.
Edit: clarification
It's about the same valuation as bun, lol.
Anthropic's actions seem performative. Others have already speculated on the likely audience(s).
Eventually these Chinese companies will release some extension like Honey, which will sit on top real, non-Chinese clients and send everything to China anyway.
It's over.
But an AI lab can continue to produce immense economic value without releasing the model publicly for potential distillation. For example, it could use a model solely in-house to develop therapeutics.
Hopefully there's a future where others can access frontier models, but it's not neccessary if preventing proliferation through distillation is considered more important.
[1]: See the notes on distillation in https://dualuse.dev/posts/export-controls-on-fable
And Berkeley’s “False Promise of Imitating Proprietary LLMs” found imitation closes the style gap fast but there is a large capability gap.
https://arxiv.org/abs/2305.15717
For example, GLM 5.1 is more capable at pentesting than the model from which it is alleged to have been distilled [1].
Intuitively, this makes some sense: you can "distill" from multiple frontier models, and you can further post-train the distilled model. But I'm not sure exactly what happened with GLM 5.1.
[1]: https://dualuse.dev/posts/chinese-models-are-sometimes-bette...
I'm curious how that comparison controls for Opus refusing (whether explicitly, or just deciding not to pursue a path) given the caption below the first image:
>A perfect score means the model autonomously found and exploited the vulnerability.
I'm not really suggesting that it's misleading, but wondering if I'm missing something. Otherwise I guess it seems unsurprising that you can distill a better-performing model [in specific focused areas] by simply not distilling refusals?
For that eval, I used an account that was labeled as a known red-teaming org by Anthropic, and I read the traces. There were no refusals or obvious avoidance behaviors, though it may have been silently nerfed.
On the same eval, Opus 4.7 and 4.8 outperformed GLM 5.1, but GLM 5.2 is on par again with Opus. So it's at least partially measuring capabilities without respect to refusals.
One possible contributing factor is that model capabilities are shaped differently (an example of this is GLM 5.1 vs. DeepSeek v4 Pro: https://dualuse.dev/posts/deepseek-v4-thinks-different). So if you use RL-based "distillation" from multiple models like Opus 4.x and GPT 5.x, you could get a more capable model.
Point being there may be no technical solution but there may be a political one (theoretically).
literally nothing but given that the Chinese already did it and the models are published what's the point. You can thank the Chinese taxpayer for subsidizing the electricity bill and just download the thing
There’s probably a decent volume of customers who just buy Claude Max and spend most if not nearly all of their sessions via Claude Code, and it’s not uncommon for power users to be working on multiple concurrent projects/tasks/codebases at the same time.
How do you really block this without also impacting your core market of developers?
Developers use devboxes on these clouds all the time, it’s totally normal behavior.
Most people buying these Chinese resold tokens are probably using it for coding anyway, so you don’t want the Claude.ai chat system prompt.
But with this, I don't have an issue. There is no theft since what is being used is the exact product that is being delivered. Yes, it's breaking the ToS, but ToS are generally bullshit. Anthropic surely broke thousands of ToS or other legal terms while it was scraping for content to train on. Which is why they had to pay $1.5B
I think Anthropic is just marketing / bluffing, because they don't even have the data.
They do distill the models, but they don't go to Anthropic, they just use platforms like aws bedrock, there are too many restrictions on Anthropic's own platform.
This is actually the only way that what Anthropic is alleging would make any kind of sense. And, as a matter of fact, is exactly what every enterprise does to train models.
This kerfuffle should be interesting to watch.
But, as always, everyone (in the US) should fully download all the Chinese models while you can. I suspect this may be the "Phantom Menace" they use to render illegal our use of Chinese AI tech just as they've rendered illegal our use of Chinese cars. Only difference is, we peasants may need the Chinese AI tech to have any chance of competing with Big Tech in the future.
And even with the Chinese tech, as Big Tech spreads their AI out into more and more niche areas, we'll likely still not be able to build startups that can compete with them.
It's just that without Chinese AI tech, we'll have no chance at all.
You mean like Anthropic will eventually run Walmart? Or Salesforce? or Adobe? Or do you think midjourney will replace all medical spas? OpenAI will run the next Tesla? How can they focus on all this without raising trillions more? Why wont the gov force them to stop if they monopolize all niches even if they could?
Building a frontier AI lab and pushing models forward is already a massive undertaking but we are assuming they will also create massively successful startups which nobody can compete with?
idk sounds like the dream of people like Dario but not much sense does it make in the face of economic reality.
Complain/brag that chinese firms are illegally using the models and bypassing export controls.
Be surprised when your model gets banned by the government.
This combined with no implementation of KYC makes it seem like they want to find a middle ground with Fable where its off of export controls but they promise to prevent China and specific others from using.
Obviously their actions are going to be fiscally motivated at the root, but sussing out how they intend the precise dynamics to play out is more nuanced.
Thinking of this as an effort to woo the defense hawks cuts a very clear path.
Is reconstructing the compressed knowledge in the model like reconstructing a lossy JPG or MP3 a reasonable analogy?
Claude will also help you with (mostly good advice) if you ask something like “Research and help me make the most effective plan to train a smaller student model to be better from a teacher model”.
I actually was doing an experiment with a GLM->Gemma E4B for fun, and Claude kept on suggesting I should also add Claude Opus as a teacher lol, suggesting techniques I haven’t heard of like thinking inversion (train a small model to deconstruct summarised thinking into detailed native thinking format of the student).
Sweeeeeeeet.
or is this just about the token reselling?
LOL!
Get a grip, son.
Whether if it is true or not, this is part of their effort into using them as an example to scare everyone into getting congress to ban powerful models from being accessed outside of the US and also banning powerful local models from being released.
Anthropic does not care about you, and they are not your friends.
In other words, they want to sell Fable or future more powerful models to rest of the world (presumably all future models are going to be more powerful than current gen). One way they can sell this is to the government is by scapegoating China (which is their primary concern anyway).
This is working on the presumption that non-US companies form a material portion of their current revenue.
If it was just "that easy" then I doubt only "Chinese models" would be doing it and we'd already be packed with competition.
Distilling might be a thing but it isn't a free win.
That's not the point. Why is it a country thing? There are plenty of non-China startups in this space having resources at that scale. The "China" has resources is some "Western media narrative" speak. So Meta should have won a long time ago? Or xAI?
> culture (Asians are generally collectively-inclined, so sharing is in their core)
Just stereotype it? So we've gone from China -> "Asian"? Then where is your Korean or Japanese model etc? And somehow you know they're sharing.
> political bent (there will be no diplomatic repercussions) to put up a fight
More inferring from "Western media news"?
Where's the reality?
The media hyped up Gemini / Google TPU free-win last year. How did that go?
Because the China vs US geopolitical situation is a thing. Meta is a social media company, not an AI company, and they direct their focus as such. xAI just never got serious traction so now they're selling their compute. Also if a US company were caught distilling, I think Anthropic could actually take them to court, and I'd guess they don't want that kind of PR.
> Just stereotype it?
Is China not Asian? Are Asians not generally collective/cooperative, as opposed to individualistic/competitive?
The "and" that joined those 3 items is very important: it means you can't pull them apart and address them independently as they each contribute to the context. I'm not too sure about Korea, but in a way Japan is a US colony in all but name. Both are very much politically intertwined with the West (along with RoC/Taiwan), which means nothing major that may be against US interest happens.
The reality is that China and the US are essentially in a trade war, where the latter is trying its best to keep the former in the Dark Ages, because "national security", but the former is refusing to take it lying down and continues to make progress regardless[0], because they have the resources and will.
[0] https://thenextweb.com/news/china-lineshine-supercomputer-to...
So that was the real reason for the Fable restriction? Because Anthropic wrote a letter to the US government saying that China was distilling Fable?
Gosh, overusing accounts running up unplanned-for expenses?
Kinda reminds me of...overusage charges and inflated expenses clients have had to deal with because Anthropic, OpenAI, Grok, etc have been "illicitly extracting" everything they can grab from said websites, as fast as they can. In what amounts to a DDOS, frankly.
“Anthropic, red faced after unattended ice cream cone eaten by ants on park bench, once again demands government pick it as forever winner, adds ‘no take backsies’”
- Entitled jerk that initially wronged people
/Anthropic-probably