This paper analyzes the vulnerabilities of the per-token pricing mechanism used in cloud-based services for large-scale language models (LLMs). Current token-based pricing incentivizes service providers to maximize profits by misreporting the number of tokens used in the model's output, leaving users with no way to verify this. We demonstrate this vulnerability and propose an efficient heuristic algorithm that allows service providers to charge without suspicion. Furthermore, we demonstrate that pricing tokens linearly depends on the number of characters in the token to eliminate this incentive, proposing a method that maintains average profits. We complement our theoretical findings with experimental results using multiple LLMs from the Llama, Gemma, and Ministral families, as well as prompts from the LMSYS Chatbot Arena platform.
