Freemium: The 98.7% Token Cut Nobody Saw Coming

Hello!

Welcome to today’s edition of Business Analytics Review!

Picture an agent that needs to move a handful of rows from Google Drive into Salesforce. A simple task. Yet under the hood, that agent was recently measured burning through 150,000 tokens just to get it done, and it still had to be told, step by step, what to do next. Rewrite the same task as a short program instead of a chain of tool calls, and the same job finishes in roughly 2,000 tokens. That is not a rounding error. That is a 98.7% reduction in token usage, and it is forcing a rethink of how agentic systems are built from the ground up.

The paradox is that nobody planned this badly. Every individual design choice behind traditional tool calling made sense in isolation: expose every capability up front so the model always has options, and route every intermediate result back through the model so it stays in full control. It is only when you stack these choices across dozens of tools and multi-step workflows that the costs compound into something absurd. Teams weren’t building inefficient agents on purpose. They were building agents the only way the original MCP pattern allowed, and the bill only showed up once they scaled.

That is the real story behind our core focus this week: how agents use context more efficiently by loading tools on demand, instead of front-loading everything and hoping the model sorts it out.