Announcement_paper_token_pricing

Is your LLM overcharging you? In a newly released paper, we look at delegated text generation in LLMs from a game-theoretic point of view and show that tokenization has important implications when it comes to pricing models. We demonstrate that the current widely used scheme, which charges users per generated token, incentivizes providers of models to manipulate the information reported to the user. We further conduct experiments with multiple model families to demonstrate the feasibility of this strategic behavior and propose an alternative pricing mechanism that completely eliminates it. Our pre-print can be found here.