LLM Pricing and Default Reasoning Allocation in Stackelberg Game
An arXiv paper (2608.13315) explores the economics surrounding large language model (LLM) services, specifically examining how providers determine per-token pricing and default reasoning-token distributions. The authors frame the interaction between a provider and a user as a Stackelberg game, where users can either accept the standard allocation, modify it, or discontinue use. While larger reasoning allocations enhance accuracy, they also lead to increased costs and latency. The paper presents a closed-form solution for the user’s optimal customized allocation and indicates that acceptable defaults either form an empty set or a compact interval for any given price. The provider’s optimal default is characterized by a three-regime rule, simplifying equilibrium computation to one-dimensional price optimization, and establishing equilibrium existence. Defaults only influence reasoning allocation when users prefer the ease of default settings; otherwise, all service outcomes align with the user’s optimal choice. Although experiments with two compact models are referenced, specifics are limited. The findings offer valuable insights into strategically designing LLM services to optimize provider revenue while catering to user needs.
Key facts
- The paper is arXiv:2608.13315.
- It models LLM service pricing as a Stackelberg game.
- Provider chooses per-token price and default reasoning-token allocation.
- User can accept default, customize, or exit.
- Larger allocations improve accuracy but increase cost and latency.
- User's optimal customized allocation is derived in closed form.
- Acceptable defaults form an empty set or compact interval for any price.
- Provider's optimal default follows a three-regime rule.
- Equilibrium computation reduces to one-dimensional price optimization.
- Defaults only matter when users value convenience of avoiding customization.
Entities
Institutions
- arXiv