Claude Code usage cost: estimates, real bills and a cost history

Claude Code usage cost is shown in two very different ways: as an estimate on your machine and as a charge in your account. For a cost history you can scroll back through, the easiest route for one developer is UsageAtlas, a free, open-source desktop app that records estimated cost per day next to tokens and models. For the money you actually paid, the bill in your Anthropic or cloud account is the only authority.

The number in /usage is a calculation

Anthropic's cost documentation is clear on this. The dollar figure in the Session block of /usage is computed locally from token counts at list price. An organization can set a modelPricing table in managed settings so the figure reflects contracted rates, but it stays an estimate either way. For Pro and Max subscribers, whose usage is part of the plan, Anthropic says the session cost figure is not relevant for billing at all.

That figure also covers one session. The plan breakdown on paid plans reaches back 24 hours or 7 days, and only on the current machine.

Where real cost lives, by setup

On every setup, OpenTelemetry can stream per-user token and cost metrics into an observability stack, which is how larger teams build their own cost history.

Why individuals still lack a cost history

Most of these pages answer "what was I charged" rather than "what did each day of coding cost, and on which model". The local files that could answer the second question are temporary: Claude Code deletes session transcripts older than cleanupPeriodDays, 30 days by default.

UsageAtlas fills that gap on the desktop. It writes daily aggregates as you work and lets you open any day, range or last month with tokens, requests, models and an estimated cost. Filters let you compare Claude Code against Codex, Cursor or Copilot on the same days, and weekday and time-of-day insights show when the expensive sessions happen. It needs no account, does not read prompts or responses, and runs on Windows, macOS and Linux.

Its cost figures are estimates too. UsageAtlas says so plainly and points to the provider account for anything that touches billing. The value is in the trend: a daily estimate that is consistent from one week to the next shows whether your spend is rising, which model drives it and which days carry it, even when the absolute figure differs from the invoice.

Which tool for which question

• "Is this session getting expensive?" /usage.
• "What did last month cost me, day by day?" UsageAtlas History.
• "I need the numbers in a script." The open-source CLI ccusage prints daily and monthly cost tables from local files, with JSON output, for as far back as the transcripts reach.
• "What will appear on the invoice?" Your billing page.

The takeaway

Treat every in-tool dollar figure as a planning number. Keep a running daily record so you can see trends across weeks, and settle any billing question in the account that charges you. For developers who switch between several AI coding tools, a single local history is the cheapest way to see where the spend goes.