OpenRouter measures the token cost of MCP, CLI tools and skills

MCP sends tool definitions on every request

OpenRouter describes the three approaches by what the model must read before acting. An MCP server sends every tool’s name, description and JSON schema on every request. A CLI presents a shell tool and a line saying the CLI exists, while a skill gives the model a one-line description and loads its full SKILL.md instructions when needed.

The company measured the GitHub MCP server, the gh CLI and a gh skill using Claude Haiku 5.5, Claude Sonnet 5.5, GPT-6 Luna and Gemini 3.8 Flash. On Claude Sonnet 5.5, the GitHub server’s 46 default tools added 19,436 tokens to every request. The same tools added 8,903 tokens on GPT-6 Luna and 8,525 through Google Vertex for Gemini 3.8 Flash. Through Google AI Studio, the Gemini figure was 13,010 tokens.

The GitHub task widened the gap

The test asked each model for the latest release of the cli/cli repository and the three most recently opened issues carrying the bug label. OpenRouter ran each combination three times on October 20, 2026, and reported median results.

On Sonnet 5.5, the CLI used 1,421 input tokens and cost $0.0075 per task. The skill used 3,315 tokens and cost $0.0118. MCP used 23,228 tokens with the two required tools, 59,417 with the 46 default tools, and 101,351 with all 91 tools. Their corresponding costs were $0.0552, $0.1274 and $0.2113. Every run in this table answered correctly.

OpenRouter says the default MCP configuration used 42 times the CLI’s input tokens and cost 17 times as much on Sonnet 5.5. The skill cost more than the bare CLI because its instructions remain in the conversation after being read. For a CLI the model already knows, the company says a skill is optional. For a user’s own CLI, it provides instructions for learning the commands.

Tool results can be as expensive as tool definitions

The comparison also examined the size of returned data. GitHub’s full release object from get_latest_release contained 42,182 characters, with 87% describing 22 download files. OpenRouter says that result added about 20,400 tokens on Sonnet 5.5, even though the task needed only the release tag and date.

OpenRouter identifies two ways to reduce the cost: expose only the tools an agent needs and shrink results before they reach the model. Scoping the server to two tools reduced the first request from 19,487 tokens to 1,406. Capping results at 4,000 characters reduced the full task to 5,357 tokens, with every run still correct in that test. The company cautions that this worked because GitHub placed the needed fields near the top of the response.

OpenRouter also tested its Tool Search server tool. It cut the upfront cost of 46 tools to under 600 tokens and halved the whole-task cost on Sonnet 5.5, while raising it on Claude Haiku 5.5. The findings give developers a practical reason to measure tool overhead on their chosen model and provider, rather than assuming that an MCP integration has a fixed cost.