One aspect that seems to be missing is the abuse of the free credits provided for new companies by AWS, Azure, and other providers.
I know of a friend's company in India who purchased inference, at 4% of the actual price and states that it gave him an unbeatable competitive edge in their large running video influence pipelines. Any new competitors could not offer their pricing at all.
Primarily that operated because registering a new company getting free AWS credits was a very tiny cost
The real problem is subscription models. Businesses want recurring revenue so they try to game the ratio of fixed subscription prices to COGS but it's always a game and so whoever can figure out the upside for the company can figure out the complementary upside for themselves.
How would one even word a bulletproof subscription contract for agentic tokens, anyway? You can't forbid automation because sub-agents are automation. You could forbid "using tokens for the benefit of more than the human who signed up" but then what do families (especially with kids) need to do? What if your friend asks you a question and you turn to a chat model? Forbidding "reselling" tokens outside of a household sounds like the closest terms but that's leaky for anyone who travels a lot, etc.
> How would one even word a bulletproof subscription contract for agentic tokens, anyway? You can't forbid automation because sub-agents are automation. You could forbid "using tokens for the benefit of more than the human who signed up" but then what do families (especially with kids) need to do? What if your friend asks you a question and you turn to a chat model? Forbidding "reselling" tokens outside of a household sounds like the closest terms but that's leaky for anyone who travels a lot, etc.
reverse the pricing structure; give modest discount once you go over certain amount of tokens, then you are incentivized NOT to start multiple accounts.
require first few transactions to be pre-paid to get around at least some of the card problems.
Of course, that would fuck over subsidized plans, but I don't see any option to keep them if you want to avoid the flood
Nice research and structuring into 4-tier layer. For providers like Anthropic and OpenAI, subscription is the entry point for all these, right? Besides the measures proposed in the article, can token usage % determine these clusters of accounts?
> For example, one operator’s price-comparison site listed a package that bought the equivalent of $3,333 worth of official Anthropic credit for 425 RMB — roughly $0.13 of usage per $1 spent.
Do these numbers make sense? $0.13 usage per $1 spent?
Yeah, I should probably clean this up. The sentence is a bit hard to understand. What I was trying to show was the steep discounts offered by resellers.
If you are using a stolen credit card to buy tokens and resell them, then the cost per token is the amount the credit card cost you (and building/running the proxy service), not the value of the tokens themselves.
thats one of the reasons why we vest any of our new customers. We need to know you before you are allowed to use our agent system. When you have an open sign up with some free credits, all hell breaks loose.
I use both of subscription and API services. on last month, i chat with CLI and let it to do something. After that, maybe in one days pass, i received the $32 USD bill. it cause my left my API key and CLI call the API to do job not through subscription.
"Token reseller market" is a fancy way of saying credit card fraud. If someone stole xboxs from stores using stolen credit cards and then sold them at 10% of their price, at what point is it a "resller market" and not "criminal enterpirse"?
I don't know anything about tokens. Does the following argument make sense?
1. Tokens are model-specific: e.g. tokens used by Anthropic cannot be used in models of other companies.
2. Tokens are generated by GPU cards. They measure the power of GPU cards.
3. Tokens cannot be separated from the models. You sort of "connect" the software part (models) into the hardware part (GPU cards) to use the tokens generated from the hardware.
Tokens measure "how much work the model did" in the same way that step counts measure "how far the person went"
GPUs "generate tokens" in the same sense that human feet "generate steps"
You can't compare token counts across different providers to get an absolute measure of "total work done" for the same reason that you can't compare step counts across different people to get an absolute measure of "total distance traveled"
I know of a friend's company in India who purchased inference, at 4% of the actual price and states that it gave him an unbeatable competitive edge in their large running video influence pipelines. Any new competitors could not offer their pricing at all.
Primarily that operated because registering a new company getting free AWS credits was a very tiny cost
How would one even word a bulletproof subscription contract for agentic tokens, anyway? You can't forbid automation because sub-agents are automation. You could forbid "using tokens for the benefit of more than the human who signed up" but then what do families (especially with kids) need to do? What if your friend asks you a question and you turn to a chat model? Forbidding "reselling" tokens outside of a household sounds like the closest terms but that's leaky for anyone who travels a lot, etc.
Fixed cost per token simply works.
reverse the pricing structure; give modest discount once you go over certain amount of tokens, then you are incentivized NOT to start multiple accounts.
require first few transactions to be pre-paid to get around at least some of the card problems.
Of course, that would fuck over subsidized plans, but I don't see any option to keep them if you want to avoid the flood
There are quite a few other mitigations that could be done by providers that aren't mentioned in the article.
Do these numbers make sense? $0.13 usage per $1 spent?
1. Tokens are model-specific: e.g. tokens used by Anthropic cannot be used in models of other companies.
2. Tokens are generated by GPU cards. They measure the power of GPU cards.
3. Tokens cannot be separated from the models. You sort of "connect" the software part (models) into the hardware part (GPU cards) to use the tokens generated from the hardware.
GPUs "generate tokens" in the same sense that human feet "generate steps"
You can't compare token counts across different providers to get an absolute measure of "total work done" for the same reason that you can't compare step counts across different people to get an absolute measure of "total distance traveled"
You should do a bit of reading on what a token is. The short answer is that it’s a series of 2-4 bytes of information turned into an integer.
Your comparisons are akin to asking “are amazon gift cards the same as a bunch of pesos?”