Insights — Where Your Tokens Go

Loading…

Headline Insights

Conversations — Context Spend

Each row is one Claude Code session (sessionId). Avg context approximates the running prompt size by summing input + cache_read tokens, divided by turns. Long conversations with low cache hit ratios are usually the ones burning the most context.

Project / Session Source Turns Total Tokens Avg Context / Turn Cache Hit Duration Cost

Duration is wall-clock time between first and last assistant message. Long durations on few-turn sessions usually mean the session was left open idle.

Cost by Project

CLI vs Cowork by Project

Per-Project Detail

Project (cwd) Cost CLI / Cowork Tokens Cache Hit Turns Top Tools

Cache hit ratio = cache_read / (cache_read + cache_create). Low ratios suggest workflows that fetch fresh data each time — optimization targets. Top 10 by cost shown by default; toggle "show all" to expand.

Tool Cost Breakdown

Cost by Category

All Tools — Approximate Cost Attribution

Cost per turn is split evenly across tools used in that turn (rough proxy — a turn with 3 tools attributes 1/3 to each). Top 15 by cost shown by default.

Tool Category Invocations Approx Cost Approx Tokens % of Total

MCP Servers

Server Calls Approx Cost Top Tools

Subagents

Subagent type Invocations Approx Cost

Subagents (Agent tool calls with subagent_type) run as separate API conversations. Costs here approximate parent-turn attribution; actual subagent compute appears in the subagent JSONL files and is included in the totals.

Cache ROI by Model

What you actually paid (Standard) vs hypothetical cost without prompt caching. Higher % saved = better ROI from cache reuse.

Model Cache Read Tokens Cache Write Tokens Output Tokens Standard Cost No-Cache Cost Saved % Saved

Source Coverage

Where your data was found. Multi-machine sync (planned Phase 2) will add per-hostname rows here.

SourceCostTurnsTokens (cache reads)Note

Methodology: Costs computed from raw JSONL using Anthropic published rates (Opus $5/$25, Sonnet $3/$15, Haiku $1/$5 per 1M; cache writes ×1.25, cache reads ×0.10). Totals on this page may exceed dashboard.html because every assistant turn (including autocompaction summaries and subagent compute) is counted individually here for attribution. All numbers are API-equivalent value — you actually pay $100/mo flat under Max 5×.