The Same Task Cost 44,026 Tokens on MCP — and 1,365 on a CLI
A benchmark ran one trivial query two ways. The MCP agent burned 32x more tokens than the CLI agent, and almost all of it was schema the model never used. The bill for that gap shows up only in production.
NeuroX AI · June 29, 2026
A June 2026 benchmark ran the simplest possible agent task — "what language is this repo?" — two ways. The CLI agent answered in 1,365 tokens. The same task through an MCP server cost 44,026 tokens, a 32x gap. The agent's actual reasoning was nearly identical. The difference was schema.
Here's the mechanism. LLM APIs are stateless, so the runtime re-sends every tool's name and JSON schema on every single turn. Load a GitHub MCP server with 40-plus tools and you inject 10–15 KB of definitions into each request — even when the agent calls one tool and ignores the other 39. It doesn't error. It just quietly pays full freight on tokens it sends a thousand times a day.
It compounds fast. The same benchmark found that loading three common servers — GitHub, Slack, Sentry — consumed 72% of a 200k context window before the agent did any work at all. The model starts every task already three-quarters full of tool definitions it mostly won't touch.
This is a prototype-to-production trap. A demo with one MCP server feels clean. At 10,000 operations a day, that schema overhead is the difference between a $120 monthly bill and a $51,000 one. The fix isn't dropping MCP — it's treating tool surface area as a budget: load lean, prefer CLI for high-frequency calls, and measure tokens-on-idle like an SLO.