Pricing and Default Token Design for LLM Reasoning Services

Summary & Methodology Analysis

This paper addresses the joint service-design problem of determining per-token prices and default reasoning-token allocations for an LLM reasoning service. Users interacting with the service can choose to accept the system default, customize the token allocation, or exit entirely. The authors formulate this provider-user interaction as a Stackelberg game, where the provider acts as the leader and the user acts as the follower. The methodology adopts standard existing mechanics, including continuous reasoning allocations, decoding-time budget forcing, and baseline probability of a correct response, expected service latency, and expected billed tokens. Furthermore, the work derives the user unique optimal customized reasoning allocation and customization value in closed form, and characterizes the user default-acceptance region and its closed-form boundaries using the Lambert W function. To compute equilibria efficiently, the authors characterize the provider optimal default through a three-regime solution, reducing equilibrium computation to a one-dimensional price optimization and a scalar comparison.

To evaluate the framework, the study uses specific models and datasets, namely Qwen3-8B and DeepSeek-R1-Distill-Llama-8B, tested across five benchmarks: AIME 2024, AIME 2025, GPQA Diamond, GSM8K, and HMMT 2025. The experimental results support the accuracy-token model and demonstrate how specific characteristics determine equilibrium prices, defaults, and allocations. The paper does not specify precise hardware setups, latency measurements in milliseconds, or dollar currency values, relying instead on the structural economic and mathematical formulations of the Stackelberg game.

Despite its mathematical rigor, the approach has notable limitations. The complete-information, representative-user model abstracts away user and task heterogeneity, private valuations, and repeated interactions. The paper does not specify how private valuations or heterogeneous user behaviors might shift the Stackelberg equilibrium in production environments with unpredictable traffic patterns. Software engineers implementing service pricing must consider that real-world workloads involve dynamic user characteristics that this simplified representative model abstracts away.