Claude Code records what every session costs, split by model. Usage Insights for Claude Code adds it up per day and per model, next to the tokens behind it and the capabilities that used them.
Cost per day
Cost per day shows what the sessions of each day cost, with the total for the range beside the title. Point at a column for that day's figure.
On a subscription plan this is what the same usage would cost at API prices, not what you pay. It is still the best measure of how much you used.
Cost by model
Cost by model splits the money by model, each in its family's colour everywhere on the page: Opus, Sonnet, Haiku and Fable. When your history has them, small tiles under it add thinking tokens (spent before answering), web searches the models made, and the subagent share: how much of all tokens your subagents used.
Tokens and the cache
Tokens & cost splits every token into input, output, cache read and cache write. Cache efficiency is the share of all input that came from the prompt cache. Higher is cheaper and faster: long sessions in one project usually score well, many short ones score low. Under it, Models counts responses per model.
Tools, connectors, skills and subagents
The row of capabilities shows what Claude used on your behalf:
- Tools used: the built-in tools, such as Bash, Read, Edit and web search, top ten by calls
- Connectors: every MCP server you called, with its number of calls
- Skills: each skill with who started it, you by name or Claude on its own, and the slash commands you ran
- Subagents / research: runs per agent type, with the average time and tool calls per run
Estimated time saved
Estimated time saved counts sessions that ended in an output, a commit or a kept edit as tasks, simple or agentic, and multiplies them by a baseline in minutes. Change the two baselines to match your work, and the range updates. It is an estimate, and the panel says how it is made.