Claude Code rolling 5-hour usage: three ledgers, three answers, and /usage is only one of them

It is a 5-hour sliding window. The clock starts when you send your first prompt, and at any moment the bucket measures the cost of all your activity in the last 5 hours. The bucket fills based on a server-side weighting (per model, peak-hour multiplier, attachments, tool calls), not on raw token counts. When the bucket crosses 1.0, every further prompt from your org returns 429 until the earliest unexpired message ages out. Per Anthropic's help docs, claude.ai chat and Claude Code share this same bucket on Pro and Max plans.

Type /usage at the Claude Code prompt. It is a snapshot of session percentage, weekly percentage, and any extra-usage balance. It interrupts whatever you were doing, prints once, and does not update. The same Help Center note that introduces /status says nothing about how to read the rolling window itself; for a live always-visible meter you need a tool that polls the same endpoint /usage is reading, but does it from outside the CLI so it does not block your loop.

The OAuth-authenticated endpoint at api.anthropic.com/api/oauth/usage. The cookie-authenticated equivalent that claude.ai/settings/usage uses is GET claude.ai/api/organizations/{org_uuid}/usage. Both return a JSON object with a five_hour field that has utilization (a float between 0 and 1, or sometimes 0 and 100 in the same payload) and resets_at (an ISO timestamp). Same shape, two auth paths. Pick whichever matches the credential you already have.

Different ledgers. ccusage walks ~/.claude/projects/*.jsonl and totals input_tokens + output_tokens across every session on this machine. The 429 is decided by the server-side five_hour utilization, which factors in browser-chat usage, per-model weights (Opus is heavier than Sonnet on the same byte count), the peak-hour multiplier Anthropic announced in late 2025, attachments, and tool-call overhead. The two numbers can drift by 30 to 40 points and still both be correct readings of what they actually measure.

Yes. Anthropic's help article 'Using Claude Code with your Pro or Max plan' is explicit: claude.ai chat and Claude Code share the same usage pool on subscription plans. If you draft a PR description in the chat tab and then run an agentic loop in Claude Code, both deplete the same five_hour bucket. Local-token tools like ccusage and Claude-Code-Usage-Monitor only see the Claude Code half because the JSONL files do not exist for browser sessions.

Yes, but not from inside Claude Code. The CLI does not have a passive status line that updates while you work. Three ways out: (1) watch claude.ai/settings/usage in a tab, (2) install a local-token monitor like ccusage --watch (sees only your Claude Code traffic, will lag the server), (3) install a server-quota monitor that polls the same endpoint /usage reads but from outside the CLI. ClaudeMeter is the third path: it polls /api/organizations/{org}/usage every 60 seconds (POLL_INTERVAL at src/bin/menubar.rs:18, POLL_MINUTES at extension/background.js:3) and renders the float in the macOS menu bar so you can glance at it during the loop.

Per-account, more precisely per-organization. The five_hour bucket is computed on the server and keyed off your org_uuid. If you run Claude Code on a laptop and a desktop logged into the same Anthropic account, both deplete the same bucket. If you also chat in a browser, that adds to the same bucket. ccusage on the laptop will only ever see the laptop's local Claude Code traffic, so on a multi-machine or chat-heavy day it underestimates the rolling 5-hour usage by a lot.

Per third-party walkthroughs (the official help docs are sparse): three numbers in one block, current session percentage, current week percentage, and any extra-usage balance from Anthropic's metered fall-through. It does not show a countdown to the rolling 5-hour reset and it does not stream. You read it once and go back to coding. If you want the resets_at timestamp the server actually returns, you have to read it from the underlying endpoint yourself or use a tool that surfaces it.

It is your own session calling the same endpoint claude.ai/settings/usage already calls every time you open the page. ClaudeMeter polls once per minute per browser, well below normal browsing traffic. The endpoint is undocumented, which means Anthropic can rename fields, not that they prohibit reading them. The Rust types in src/models.rs declare every Window field as Option, so when fields change the next brew release patches them in one line.

Run a local-token tool (ccusage) and a server-quota tool (ClaudeMeter, or ccusage with the OAuth flag) at the same time. The first answers 'what did this machine spend?'. The second answers 'when does my next prompt 429?'. Treating them as alternatives is a category error; they read non-overlapping sources and give you complementary data. The /usage slash command is fine as a one-shot sanity check but it will not save you from a mid-refactor 429 because by the time you remember to type it, you are already at 91 percent.