Analysis of Benchmark Performance, Token Price, and the Market Share of Large Language Models

Abstract

The growing number of large language models (LLMs) makes it important to understand the factors associated with their adoption. Although token price and benchmark performance are commonly used to compare models, their relationship with model market share has not been established. Our study examines whether listed token price and SWE-rebench performance scores are associated with model market share. We analyzed 202 coding-capable models available on OpenRouter over an eight-week period following entry into the platform’s weekly top-50 rankings. Model market share was measured as the share of total weekly token usage. Most models did not gain share after entry, with the median model declining to roughly half its entry share. Correlation analysis indicated that entry token share was positively associated with subsequent model market share (Spearman’s ρ = 0.615), with larger entrants declining faster but retaining the largest shares after eight weeks. Regression analysis found that neither SWE-rebench performance nor listed token price was significantly associated with subsequent share after controlling for entry token share. These findings suggest that early platform traction is more strongly associated with subsequent market share than either benchmark performance or price.