I Built a Token Cost Analyzer Skill for Claude Code. Here’s What I Found

In this post:The gap in my dashboardWhy this matters right nowA quick cache primerThe buildWhat the numbers taught me about token efficiencyOpus 4.7 Updated SkillWhat’s nextThe gap in my dashboardI’ve been building with Claude Code daily for weeks. Videos, MCP servers, skills, GUI tools. I already track costs through Claude’s OpenTelemetry integration piped into Grafana. The aggregate picture looked good: 97.8% cache hit ratio, $2,171 in estimated savings across 44 sessions.Thanks for reading! Subscribe for free to receive new posts and support my work.But aggregates only tell you what happened in total across all Claude Code projects. They don’t tell you what happened at turn 19 (chat response) within an individual Claude Code chat session.Cost forecasts, cache hit rates, model efficiency, token usage over time. All useful at the macro level. But none of it shows me why one response costs $0.92 and the next costs $0.03, or where the cache breaks inside a conversation, or what happens…