I wrote about a practical approach to converting AI token usage into credits without floating-point arithmetic. The key insight is using multi-layered integer scaling (price scaling × credit conversion × per-million-token base) to maintain precision across millions of transactions.
I hate the current token gym membership pricing models, they are too complicated and abstract and even worse, you have issues like what this post points out. They favor the providers, not the users. They create a lot of stress for users who have to decide if they can afford to point GasTown at their codebase.
The problem is that the companies serving LLMs can't delineate compute between users. You're using a cloud of compute, not a single individual unit of compute. If you could measure at the unit of compute, then you'd just pay by the minute for what you use instead of by the token, we would get out of this situation.
Yes, I agree with your points. When it comes to using compute minutes instead of an I/O credit system, I do not think providers will adopt that model. Without support from them, application developers cannot implement it this way, because it becomes more costly when applications and providers use different approaches.
I wrote about a practical approach to converting AI token usage into credits without floating-point arithmetic. The key insight is using multi-layered integer scaling (price scaling × credit conversion × per-million-token base) to maintain precision across millions of transactions.
I work in the payments industry. Just do everything in cents and divide by 100. Makes your life a lot easier.
I hate the current token gym membership pricing models, they are too complicated and abstract and even worse, you have issues like what this post points out. They favor the providers, not the users. They create a lot of stress for users who have to decide if they can afford to point GasTown at their codebase.
The problem is that the companies serving LLMs can't delineate compute between users. You're using a cloud of compute, not a single individual unit of compute. If you could measure at the unit of compute, then you'd just pay by the minute for what you use instead of by the token, we would get out of this situation.
Yes, I agree with your points. When it comes to using compute minutes instead of an I/O credit system, I do not think providers will adopt that model. Without support from them, application developers cannot implement it this way, because it becomes more costly when applications and providers use different approaches.