Fresh-Session Token Baseline (31K)

Description

Measured token consumption of a Claude Code session before any user work: ~31K tokens composed of system prompt (4.3K), built-in system tools (20.3K for ~40 tools), MCP tools (~905), custom agents (1.6K), memory files (1.6K), and skills (2.3K). Documented in GitHub issue #24243. Establishes the floor against which per-turn costs must be measured and makes visible why the effective window is smaller than the nominal 200K.

Key claims

Relations

Sources

src-20260409-e32facfa1c9f