Claude Code: 86% of token costs hide in invisible overhead, not your prompts

Claude Code's cost structure is opaque by design. Every turn includes the full conversation history, all MCP server schemas, and the system prompt—but none of this is visible in the Anthropic Console or the /cost command. A trace of real usage showed that prior assistant context alone (replies and tool outputs replayed from earlier turns) consumed 30–45% of input spend, yet most developers assume costs are driven by their prompts and the code Claude generates. The problem compounds at scale: Uber deployed Claude Code to roughly 5,000 engineers and saw bills climb to $500–$2,000 per month per person, while a single developer on a $200 Max plan burned $50,000 in one month—numbers that suggest widespread configuration drift rather than model cost alone.

The root cause is that configuration decisions accumulate invisibly. A developer adds an MCP server for one project and never removes it; CLAUDE.md files grow to 5,000–10,000 tokens over months; thinking effort defaults to "high" (later changed to "medium" after SWE-bench data showed 76% fewer output tokens at medium effort with the same completion rate). None of these choices show up in the org-level console. Checkmarx and other supply-chain scanning tools track API imports, but Claude Code's context is not a lockfile—it lives in configs, datasets, and session state, so security and cost teams track fragments instead of a unified view.

Comet's Cost Intelligence closes this gap by intercepting API traffic, mapping every token to its source category (prior assistant context, tool results, model choice, thinking effort, MCP schemas), and surfacing recommendations with estimated savings. By connecting attribution (which categories eat tokens) to policy (what should change) to implementation (applying changes org-wide), teams can move from blanket restrictions (which hurt productivity) to informed decisions—e.g., switching default model to Sonnet (which costs 0.6× Opus per tier and handles comparable development work) or blocking MCP servers with zero calls that still load schemas every turn.