Your AI agent might be doing the same work for 40 times the cost. That’s not an estimate. That’s what OpenRouter measured when it compared three different architectural approaches for connecting AI models to tools.
The test used a simple task on Claude Sonnet 5.5. Using MCP (Model Context Protocol, an open standard that lets AI models connect to external tools and data sources), the agent consumed 59,417 input tokens. The same task via CLI, a traditional command-line interface approach, used 1,421 tokens. A “skills” method, where tool definitions are pre-packaged more efficiently, used 3,315 tokens. Same task. Same model. Same outcome. The cost difference between MCP and CLI was roughly 41 to one.
That number deserves to land.
MCP has become the dominant standard for building AI integrations, and for good reason. It’s well-documented, widely supported by major AI providers, and easy to get started with. The issue is structural. MCP loads the complete definition of every available tool at the start of each request, sending that entire payload to the model every single time, whether the agent ends up using those tools or not. That’s a lot of overhead for a task that might only need one tool.
For a single request, the overhead is barely noticeable. Run 100,000 agent requests per month, and the math changes fast. At that volume, the token cost difference between MCP and CLI translates into a substantial gap in your monthly API bill. This is the kind of number that can quietly turn a profitable AI automation project into an expensive one.
The good news is that the problem is solvable. A technique called Tool Search defers loading tool definitions until the agent actually needs them. With Tool Search in place, MCP’s upfront token cost drops to under 600 tokens per request. That’s cheaper than the Skills approach and close to CLI efficiency. The architecture doesn’t have to cost what it costs out of the box.
The takeaway isn’t to avoid MCP. It’s widely used for good reasons, and the ecosystem around it is maturing quickly. The takeaway is that default configurations are rarely optimal configurations, and the choices your technical team makes when building an AI pipeline have direct financial consequences at scale. Most teams adopt defaults because they’re the path of least resistance. Few audit whether those defaults are still the right choice once usage grows beyond the prototype stage.
If you’re evaluating or already running AI agent workflows, the architecture question isn’t academic. At meaningful scale, a 41x token cost difference is the difference between a project that makes financial sense and one that quietly bleeds budget. That conversation belongs at the planning table, not the retrospective.
Want to explore how AI agent architecture could benefit your business? Let’s talk.