I built a pricing API for LLMs — then realized the real users might not be human

LLM Price Watch began its journey as a straightforward solution to compare token pricing across various AI models such as Claude, GPT, Gemini, DeepSeek, and Grok. Initially, the project consisted of creating a simple calculator and an associated API. However, a more significant realization emerged when the API's fourth endpoint, GET /v1/recommend, was developed.

This endpoint did more than merely provide pricing; it recommended the most suitable model for a specific use case, considering both cost and editorial analysis. Consequently, the API's primary audience shifted from "developers building cost dashboards" to "AI agents making runtime tool selections." To cater to this new audience, the API's design was reimagined to provide structured, agent-callable responses that include reasoning.

The developers decided to open CORS, eliminate the need for an API key, and avoid any friction that could hinder the API's usage in client-side agent code. Although a paid tier with rate limits was eventually planned, it was intentionally postponed to avoid hindering the API's initial adoption. Ensuring the accuracy of the pricing data was crucial; therefore, it was verified directly from each provider's official pricing page.

The API's backend, initially designed for a specific use case, unexpectedly proved to be reusable infrastructure. It became an integral part of another site, StackIndex AI, which utilized the API to provide real-time cost estimations and model recommendations. This cross-site integration underscored the API's versatility and the importance of designing responses for readers who cannot ask follow-up questions.