For this talk, I want to present an experimental benchmark exploring how poor API design increases the number of tokens AI models use when generating integration code. I compare two APIs with identical business functionality: (1) a well-designed REST/HAL API with consistent naming, predictable resource paths, and hypermedia links, and (2) a poorly designed RPC-style API with inconsistent casing, POST-for-read operations, nested payloads, abbreviations, and no hypermedia navigation. I tested frontier models across OpenAI, Anthropic, and Google families using the same coding task and measured token consumption. The results were consistent across every model: All models used more completion tokens with the bad API. Token overhead ranged from approximately 30% to 77%, the average overhead was around 52%, and GPT-5-mini additionally exposed higher reasoning-token usage, showing that the models literally had to “think harder” when consuming poorly designed APIs. In the talk, I’d show that the largest source of inefficiency is not just inconsistent naming but RPC-style workflows, POST requests for reads, missing hypermedia navigation, and APIs that force models to infer relationships instead of following explicit links. The talk argues that API design now directly impacts AI inference cost, code generation speed, generated code complexity, and reliability of AI-assisted integrations.