People trading their unused credits feels more genuine, although still in violation of the agreements. The person who got into YC Startup School who was trying to resell the $2500 of credits was interesting. It wouldn’t be that hard for OpenAI to identify the IP addresses of the relays and start flagging accounts, tracing it back to the source. Risking burning your bridges with YC for a relatively small profit is a questionable decision.
The original article showed discounts ranging all the way up to 98%. At those levels it’s obviously not people reselling anything. It’s either sourced from stolen API keys, bought with stolen credit cards, or acquired through automated sign up of trial accounts if you’re actually getting the API you request.
I would expect a lot of them are reselling a different API. Sign up for Anthropic tokens and get Deepseek responses instead.
Generally speaking, B2B prices are rarely supply-and-demand priced in the usual sense.
YC has advised startups in the past that it's easier to sell a single $100k customer than 100 $1k customers.
It would also be relatively surprising to learn that i.e. the Chinese providers are OOMs better at inference than OAI/Anthropic (like their prices would imply if they were in a perfectly competitive market).
Fair, I hadn't looked recently. It looks like currently kimi is either 1/2 or 1/4 Ant pricing depending on whether you think Opus 5 is usable or not. (Deepseek is still an OOM though)
Costs also have to include the amortized training costs.
But the API price is likely simply regular supply and demand, charging as much as the market will pay. Corporations are dropping insane amounts because it's still peanuts for many industries. Software has just been ridiculously cheap before AI. So high prices are still low for companies if it eases some bottlenecks.
The costs only have to include amortized training costs if you are trying to be profitable overall. Having positive unit economics and VC subsidized fixed costs is pretty standard.
Wait a sec, I have to trust a third party with basically no reputation, did I get it right?
It's basically asking for being hacked and/or sending you private data to random email addresses! Neither at a 99% discount I'd do it.
I understand if someone, for any reason, cannot access a specific model ... But nowadays, there are so many alternatives that even this doesn't make sense any more.
It's really common in China where OpenAI and Anthropic models are firewalled. There was a joke that Chinese people didn't realize how cheap DeepSeek was, because they already were using resold ChatGPT/Opus tokens at that price.
Here is a more detailed article about how it works:
I expect most of the use to be for bulk data processing or desperate founders who don't care, not for agentic coding use at promising startups.
If your startup needs to run a million records of something, especially public data, through an LLM to extract the data you need, using bootleg tokens to shrink the bill starts feeling tempting.
If you're concerned about the data leaking, the biggest risk is that the API backends are quietly routing your requests to a cheaper model. You might be trying to buy Opus tokens but get Deepseek Flash responses.
> to run a million records of something, especially public data, through an LLM to extract the data you need
Maybe this make sense, but anyway I have to pay a lot of attention at the output I get. Eg: who guarantees there is no prompt/sql injection? Especially if I have to load the output in some internal system.
Bulk data processing typically has structured JSON output.
I mean someone could try to sneak prompt injection into a text field, but the people buying black market resale tokens from third parties aren’t thinking about anything other than getting cheap output.
So what? The world is more trustworthy than you can imagine. I have bought over 30 GPUs over the internet, sight unseen. From around the world and every single one has arrived good. I have also gotten over 20 from marketplace and the same. Use your common sense, from conversation and everything else, you can often separate the scammers from real folks.
I'd rather trust a third party with no reputation than a well known Misanthropic company led by a man whose wife was asking Epstein for investment. They are even less trustworthy then ClosedAI.
Distillation is one of the most unique and interesting aspects of this.
But otherwise, if a company gives something valuable for creating an account on their platform, expect that people will automate the creation of millions of accounts. If employees of B2B partners get benefits, they will resell them. Accounts will be hacked and resold. The same basic abuse patterns are decades old for online delivery services, loyalty accounts for airline and hotels, etc. There are entire industries dedicated to those spaces as well: large organizations with physical offices, hundreds of employees, HR departments, etc. dedicated to reselling digital benefits on grey markets.
Some companies are tolerant of allowing this to happen. The pessimistic view is that even illegitimate traffic contributes to the KPIs that your investors care about. The slightly less pessimistic view is that fraud prevention will always have trade-offs and false positives, and sometimes the savings of preventing fraud are genuinely outweighed by the false positives. Or maybe it's just Hanlon's razor and they truly never saw it coming.
This was specifically meant to be about the ai credit resellers, not the relays themselves. I put together another piece here: https://vectoral.com/blog/token-relay-market that tackles that side of the market.
The reseller could use an intermediate proxy and modify the traffic like in [1], to get control of the client machine - depending on the harness permissions.
TLS terminates at the proxy (say, https://reselltokens.ai), end to end integrity is not enforced. LLM traffic contains tool calls like "bash ...", which are executed on the client machine, they can be manipulated. Secret exfil is also possible.
I swear every time I think I have a unique tweak for an OSS tool, it turns out someone else did it. I forked CLIProxyAPI myself to add something similar, though my implementation is closer to a provider-egress DLP layer and it's intended for API keys and the like, not PII. It uses gitleaks but also supports reversible redaction so a secret that's flagged can be replaced with a placeholder before CLIProxyAPI sends the request to the provider and then any instances of that placeholder can be rehydrated in the response to the caller.
Since it modifies logic across the full request/response lifecycle, I unfortunately couldn't implement it cleanly with the existing plugin API.
Most of these are your standard botnet rings. Either accounts directly are taken over, and the attacker adds 2FA or carding rings take stolen #s and attempt to add credits.
It is...incredible how many there are. Stripe does far too little in my opinion to help prevent issues like this, even though they have the business intelligence and enough data to do so.
A common refrain is that oh there are such great margins on tokens that none of this matters… I wonder how long until that notion will be disavowed? The scale of the tokensnaffling is massive, not just from resale, but also people using multiple subscriptions. The amount of subsidization is only growing, every week it seems like OpenAI and Anthropic are doing “resets” which allow a single $200 subscription to incur $20k+ of usage (if billed at API rates). At some point we must all surely accept that the economics of this do not work.
Or API rates are their way of doing price discrimination and it's all profitable. Make sure Enterprises(tm) pay the Enterprise(tm) rate, but mop up the rest of the demand too at lower price points. This has been part of the software playbook for decades.
"My rough estimate is that, across the sites, forums, and resellers I looked at, there are probably tens of millions of these credits being offered." Yeah very useful statemenet it's not like everyone spends hundreds of millions of tokens per day on the 100 or 200$ plan
From talking with some of the companies experiencing this, I can confirm that a portion of it is actual credit card fraud. Tokens have become a pseudo-currency, making them a prime target for abuse.
Some of the abuse is more benign, but there is also real fraud through chargebacks, account takeovers, and stolen credit cards.
Any situation where you can extract a benefit from a stolen credit card and then sell that benefit before the chargeback renders it moot will have some fraud-based usage.
They were never offering actual US models. We tested them extensively and it was clear whatever they were passing as Anthropic were simply mediocre distillations based on Kimi.
Haha! Incredible. What a scam! Too good to be true, I suppose. I never used it for anything meaningful because everything meaningful is also tragically sensitive so I suppose I didn't notice.
yep, with projects like open api/newapi it takes maybe 10 min to pin up and openAi compatible poxy gateway. The cheap prices look tempting until the relay operator logs all your prompts or the upstream account gets banned mid-request.
What is the logic behind selling at such a steep discount? Lack of buyers or trust? Seems insane to sell tham at -50% or more, especially since you can use them yourself (if you're a startup)
Risk-reward calculus. A buyer may not take the risk or hassle of going underground only for a small upside. Basically it is not as tempting at 90% of list price, but it is at 10% .
Chinese. There are large number of Chinese people who are dependent on the Western model because they're still ahead of the game. But they're continuously getting banned and getting super frustrated.
Especially for Claude because Anthropic is very good at identifying mainland Chinese and getting them banned in hours. There are many of them who are willing to pay more than the original rate for a stable experience.
It's very hard for them because they'll need a legit phone number and bank cards that are not issued in China, and a clean enough IP, etc. and those better match together to make sense. (Back in the day, ChatGPT required resident IPs, which made it worse, but they worry about growth more now). Obviously, they have to use a VPN to access the real Internet, and most of the IPs they can find are shared with bots and abusers.
So let's see, we have the following factors in play:
1) Capitalism - Adam Smith, John Maynard Keynes (Keynesian Economics), etc., etc. in most places in the world...
2) Huge validated existing international market...
3) Multi-jurisdictional World... laws/statutory codes applicable to businesses in specific circumstances in one place may not be applicable to businesses in specific circumstances in another...
4) AI Tokens are a commodity; i.e., there is no chokepoint or monopoly controlled by one AI company in one jurisdiction, i.e., if one AI company makes rules unacceptable to a token consumer, that consumer can simply switch providers to another provider in another jurisdiction somewhere else in the world.
5) Tokens can be bought, sold, and resold at profit just like any other good or service.
6) Tokens can be bought from anywhere in the world and sold to anywhere in the world. Easily.
7) Tokens are a digital good, easy to scale, and do not require supply chains, lead times, labor, manufacturing, warehousing, shipping, going through geographic chokepoints, customs, etc., etc. -- all of the things that manufactured goods do.
8) Many people around the world want to make money or make more money... i.e., "economic incentive" (aka Capitalism's "profit motive")...
Well... add all of those together and what do you get?
You get buy/sell/trade forums/auctions/individuals/brokers/businesspeople -- around that market...
Just like you get those same things around every other market.
In this large multi-jurisdictional world, if one government makes all of that illegal in their country, then another government is going to be happily collecting all of the taxes from making all of that legal, in theirs!
If a given government makes trade illegal -- then they correspondingly lose the tax revenue...
Taxes and trade are intricately, intricately intertwined...
Could this business model be used for money laundering or other illegal activities?
Yes -- but any other business model could as well!
And, on the flip side, this business model could be accomplished legally/lawfully/morally/ethically -- just like any other business where there is an actual underlying value being exchanged.
Because, AI Tokens, if legally/lawfully/morally/ethically traded, do have underlying value...
In conclusion, at this point in time, I am neither for this business model nor against it...
But I think it'll be highly interesting to watch this space for the next couple of years, to see what happens, to see who does what, to see what plays out on the legal front, on the government front (foreign + domestic), on the media front, and on the technology front surrounding it...
The original article linked in the opening has more context https://vectoral.com/blog/token-relay-market
People trading their unused credits feels more genuine, although still in violation of the agreements. The person who got into YC Startup School who was trying to resell the $2500 of credits was interesting. It wouldn’t be that hard for OpenAI to identify the IP addresses of the relays and start flagging accounts, tracing it back to the source. Risking burning your bridges with YC for a relatively small profit is a questionable decision.
The original article showed discounts ranging all the way up to 98%. At those levels it’s obviously not people reselling anything. It’s either sourced from stolen API keys, bought with stolen credit cards, or acquired through automated sign up of trial accounts if you’re actually getting the API you request.
I would expect a lot of them are reselling a different API. Sign up for Anthropic tokens and get Deepseek responses instead.
Another part of the discount is that Claude Max 20x is $200 but gives usage equivalent to thousands of dollars worth of API-based token spend.
But also, resellers only need to make an overall profit including kickbacks from the companies purchasing token history for distillation.
It's also unclear wherever the subscription price is the real cost, or the API.
I suspect it's closer to the sub price and anthropic is just milking their API users, but that's something you'd only know from the inside
Generally speaking, B2B prices are rarely supply-and-demand priced in the usual sense.
YC has advised startups in the past that it's easier to sell a single $100k customer than 100 $1k customers.
It would also be relatively surprising to learn that i.e. the Chinese providers are OOMs better at inference than OAI/Anthropic (like their prices would imply if they were in a perfectly competitive market).
Chinese prices aren't really OOM cheaper. Deepseek recently did a big price hike too.
Fair, I hadn't looked recently. It looks like currently kimi is either 1/2 or 1/4 Ant pricing depending on whether you think Opus 5 is usable or not. (Deepseek is still an OOM though)
Costs also have to include the amortized training costs.
But the API price is likely simply regular supply and demand, charging as much as the market will pay. Corporations are dropping insane amounts because it's still peanuts for many industries. Software has just been ridiculously cheap before AI. So high prices are still low for companies if it eases some bottlenecks.
The costs only have to include amortized training costs if you are trying to be profitable overall. Having positive unit economics and VC subsidized fixed costs is pretty standard.
Open weight SOTA models are not greatly cheaper than Anthropic and providers don't have to cover training capex.
Wait a sec, I have to trust a third party with basically no reputation, did I get it right?
It's basically asking for being hacked and/or sending you private data to random email addresses! Neither at a 99% discount I'd do it.
I understand if someone, for any reason, cannot access a specific model ... But nowadays, there are so many alternatives that even this doesn't make sense any more.
It's really common in China where OpenAI and Anthropic models are firewalled. There was a joke that Chinese people didn't realize how cheap DeepSeek was, because they already were using resold ChatGPT/Opus tokens at that price.
Here is a more detailed article about how it works:
https://www.chinatalk.media/p/how-to-buy-cheap-claude-tokens...
Oh that's interesting, looks like those proxies are probably also selling the data to Chinese AI companies for training.
I expect most of the use to be for bulk data processing or desperate founders who don't care, not for agentic coding use at promising startups.
If your startup needs to run a million records of something, especially public data, through an LLM to extract the data you need, using bootleg tokens to shrink the bill starts feeling tempting.
If you're concerned about the data leaking, the biggest risk is that the API backends are quietly routing your requests to a cheaper model. You might be trying to buy Opus tokens but get Deepseek Flash responses.
> to run a million records of something, especially public data, through an LLM to extract the data you need
Maybe this make sense, but anyway I have to pay a lot of attention at the output I get. Eg: who guarantees there is no prompt/sql injection? Especially if I have to load the output in some internal system.
Bulk data processing typically has structured JSON output.
I mean someone could try to sneak prompt injection into a text field, but the people buying black market resale tokens from third parties aren’t thinking about anything other than getting cheap output.
So what? The world is more trustworthy than you can imagine. I have bought over 30 GPUs over the internet, sight unseen. From around the world and every single one has arrived good. I have also gotten over 20 from marketplace and the same. Use your common sense, from conversation and everything else, you can often separate the scammers from real folks.
I mean reselling items bought with stolen credit cards have been a thing for at least 20 years
I don't risk being hacked if I buy some stuff on vinted.
I'd rather trust a third party with no reputation than a well known Misanthropic company led by a man whose wife was asking Epstein for investment. They are even less trustworthy then ClosedAI.
Distillation is one of the most unique and interesting aspects of this.
But otherwise, if a company gives something valuable for creating an account on their platform, expect that people will automate the creation of millions of accounts. If employees of B2B partners get benefits, they will resell them. Accounts will be hacked and resold. The same basic abuse patterns are decades old for online delivery services, loyalty accounts for airline and hotels, etc. There are entire industries dedicated to those spaces as well: large organizations with physical offices, hundreds of employees, HR departments, etc. dedicated to reselling digital benefits on grey markets.
Some companies are tolerant of allowing this to happen. The pessimistic view is that even illegitimate traffic contributes to the KPIs that your investors care about. The slightly less pessimistic view is that fraud prevention will always have trade-offs and false positives, and sometimes the savings of preventing fraud are genuinely outweighed by the false positives. Or maybe it's just Hanlon's razor and they truly never saw it coming.
This research is way toooo shallow. He really should go check out linux.do or nodeseek.com — token resale economy there is truly breathtaking.
Edit: https://vectoral.com/blog/token-relay-market mentioned in comment.
This was specifically meant to be about the ai credit resellers, not the relays themselves. I put together another piece here: https://vectoral.com/blog/token-relay-market that tackles that side of the market.
The reseller could use an intermediate proxy and modify the traffic like in [1], to get control of the client machine - depending on the harness permissions.
TLS terminates at the proxy (say, https://reselltokens.ai), end to end integrity is not enforced. LLM traffic contains tool calls like "bash ...", which are executed on the client machine, they can be manipulated. Secret exfil is also possible.
[1] https://arxiv.org/html/2604.08407v1
You can place your own CLIProxyAPI instance in front of the reseller proxy and block prompts/responses that look harmful.
There are community plugins like this: https://github.com/rheodev/cpa-plugin-privacyfilter
I haven't tried the plugin system myself yet.
I swear every time I think I have a unique tweak for an OSS tool, it turns out someone else did it. I forked CLIProxyAPI myself to add something similar, though my implementation is closer to a provider-egress DLP layer and it's intended for API keys and the like, not PII. It uses gitleaks but also supports reversible redaction so a secret that's flagged can be replaced with a placeholder before CLIProxyAPI sends the request to the provider and then any instances of that placeholder can be rehydrated in the response to the caller.
Since it modifies logic across the full request/response lifecycle, I unfortunately couldn't implement it cleanly with the existing plugin API.
Nice, he can help fix my Linux Bluetooth audio issues.
that one platform is using a flipped chroma logo - chroma has nothing to do with this racket (source : i’m chroma’s ceo)
Most of these are your standard botnet rings. Either accounts directly are taken over, and the attacker adds 2FA or carding rings take stolen #s and attempt to add credits.
It is...incredible how many there are. Stripe does far too little in my opinion to help prevent issues like this, even though they have the business intelligence and enough data to do so.
A common refrain is that oh there are such great margins on tokens that none of this matters… I wonder how long until that notion will be disavowed? The scale of the tokensnaffling is massive, not just from resale, but also people using multiple subscriptions. The amount of subsidization is only growing, every week it seems like OpenAI and Anthropic are doing “resets” which allow a single $200 subscription to incur $20k+ of usage (if billed at API rates). At some point we must all surely accept that the economics of this do not work.
Or API rates are their way of doing price discrimination and it's all profitable. Make sure Enterprises(tm) pay the Enterprise(tm) rate, but mop up the rest of the demand too at lower price points. This has been part of the software playbook for decades.
"My rough estimate is that, across the sites, forums, and resellers I looked at, there are probably tens of millions of these credits being offered." Yeah very useful statemenet it's not like everyone spends hundreds of millions of tokens per day on the 100 or 200$ plan
This was poorly worded on my part; I meant in terms of dollars, not tokens.
There's a lot of assertion here and the other article that this fraud and abuse but is there any evidence of that?
A simpler explanation is that that this is just a resale market.
From talking with some of the companies experiencing this, I can confirm that a portion of it is actual credit card fraud. Tokens have become a pseudo-currency, making them a prime target for abuse.
Some of the abuse is more benign, but there is also real fraud through chargebacks, account takeovers, and stolen credit cards.
Any situation where you can extract a benefit from a stolen credit card and then sell that benefit before the chargeback renders it moot will have some fraud-based usage.
Cyberpunk dystopia vibes
> Cyberpunk dystopia vibes
right. Abstractions taken to the max. When tech solves problems that only 0.001% care about. NFT smelt similar.
Men breaching Sauron's ToS by reselling their nine rings of power, and they think they are beating the system.
> I checked where you’d expect to find underground marketplaces. Telegram had a few channels, with one being relatively active.
It's not exactly "underground" if they clearly advertising public channels out in the open.
Yunwu.ai was a nice way to get ultra-cheap tokens that someone will probably spy on you with, but they no longer offer US models.
At first I thought it was so people could steal the traces, but now I wonder if this isn't just laundering startup credits for dollars.
They were never offering actual US models. We tested them extensively and it was clear whatever they were passing as Anthropic were simply mediocre distillations based on Kimi.
Haha! Incredible. What a scam! Too good to be true, I suppose. I never used it for anything meaningful because everything meaningful is also tragically sensitive so I suppose I didn't notice.
There are like thousands sites with similar features all using newapi core.
You can easily find them in Chinese tech forum linux.do
yep, with projects like open api/newapi it takes maybe 10 min to pin up and openAi compatible poxy gateway. The cheap prices look tempting until the relay operator logs all your prompts or the upstream account gets banned mid-request.
So are they stealing from their employers or is something else going on?
nope - a.i companies to promote their platforms offer 'a.i' credits to startups usually worth 10ks, 100ks worth of credits in $ terms.
so to a startup - you can trade your credits - then get actual cash.
just like you would if trading debt etc.
What is the logic behind selling at such a steep discount? Lack of buyers or trust? Seems insane to sell tham at -50% or more, especially since you can use them yourself (if you're a startup)
Risk-reward calculus. A buyer may not take the risk or hassle of going underground only for a small upside. Basically it is not as tempting at 90% of list price, but it is at 10% .
Stolen credit cards
Demo accounts
Free trials
Unlimited chat relays (eg chatgpt chat)
Leaked company credentials
Etc
And I’m pretty sure that’s just the tip of the iceberg.
What does the demand side look like? I understand why people have these tokens they want to sell, who wants to buy them?
Chinese. There are large number of Chinese people who are dependent on the Western model because they're still ahead of the game. But they're continuously getting banned and getting super frustrated.
Especially for Claude because Anthropic is very good at identifying mainland Chinese and getting them banned in hours. There are many of them who are willing to pay more than the original rate for a stable experience.
It's very hard for them because they'll need a legit phone number and bank cards that are not issued in China, and a clean enough IP, etc. and those better match together to make sense. (Back in the day, ChatGPT required resident IPs, which made it worse, but they worry about growth more now). Obviously, they have to use a VPN to access the real Internet, and most of the IPs they can find are shared with bots and abusers.
Hey Claude, scan these marketplaces for arbitrage opportunities.
The thing will eat itself unless the AI companies find a way to make money directly from it.
This would be a fantastic method of getting data for distillation, would be surprised if that's not why the tokens are so cheap
Cool to see my thesis validated.
https://news.ycombinator.com/item?id=48664223
man seems like this entire thread invited a bunch of ads for exactly the thing talked about in the article
can't they detect if someone is reselling their tokens like this ? doesn't seem too hard
> CREDITS FROM YC STARTUP SCHOOL
Join YC, get free shit from the network, profit. Nice.
As a bonus, you’ve got a story for the “tell me about a time you hacked a non computer system” question on the YC application
> CREDITS FROM YC STARTUP SCHOOL
Is that still a thing? I’ve thought they’ve discontinued the deals section. (There are a lot of other ways to get a startup grant, of course.)
What will you pitch to YC to get accepted?
So let's see, we have the following factors in play:
1) Capitalism - Adam Smith, John Maynard Keynes (Keynesian Economics), etc., etc. in most places in the world...
2) Huge validated existing international market...
3) Multi-jurisdictional World... laws/statutory codes applicable to businesses in specific circumstances in one place may not be applicable to businesses in specific circumstances in another...
4) AI Tokens are a commodity; i.e., there is no chokepoint or monopoly controlled by one AI company in one jurisdiction, i.e., if one AI company makes rules unacceptable to a token consumer, that consumer can simply switch providers to another provider in another jurisdiction somewhere else in the world.
5) Tokens can be bought, sold, and resold at profit just like any other good or service.
6) Tokens can be bought from anywhere in the world and sold to anywhere in the world. Easily.
7) Tokens are a digital good, easy to scale, and do not require supply chains, lead times, labor, manufacturing, warehousing, shipping, going through geographic chokepoints, customs, etc., etc. -- all of the things that manufactured goods do.
8) Many people around the world want to make money or make more money... i.e., "economic incentive" (aka Capitalism's "profit motive")...
Well... add all of those together and what do you get?
You get buy/sell/trade forums/auctions/individuals/brokers/businesspeople -- around that market...
Just like you get those same things around every other market.
In this large multi-jurisdictional world, if one government makes all of that illegal in their country, then another government is going to be happily collecting all of the taxes from making all of that legal, in theirs!
If a given government makes trade illegal -- then they correspondingly lose the tax revenue...
Taxes and trade are intricately, intricately intertwined...
Could this business model be used for money laundering or other illegal activities?
Yes -- but any other business model could as well!
And, on the flip side, this business model could be accomplished legally/lawfully/morally/ethically -- just like any other business where there is an actual underlying value being exchanged.
Because, AI Tokens, if legally/lawfully/morally/ethically traded, do have underlying value...
In conclusion, at this point in time, I am neither for this business model nor against it...
But I think it'll be highly interesting to watch this space for the next couple of years, to see what happens, to see who does what, to see what plays out on the legal front, on the government front (foreign + domestic), on the media front, and on the technology front surrounding it...
When "token futures contracts", start selling. Then the bubble is about to burst.
CME is launching GPU compute futures in October: https://www.cmegroup.com/markets/energy/power/compute-future...
These are amateurs compared to what's happening in China.
Explain please?
They built this wall …