The Claude Code rate limit is eight floats, not one number

There isn't one. Claude Code on Pro and Max is gated by eight separate utilization floats returned by GET claude.ai /api/organizations/{org_uuid}/usage. Each float lives between 0.0 and 1.0; any one of them hitting 1.0 returns a 429 on the next prompt. The four visible ones on claude.ai/settings/usage are five_hour, seven_day, seven_day_sonnet, and seven_day_opus. The four the UI does not show are seven_day_oauth_apps, seven_day_omelette, seven_day_cowork, and extra_usage.utilization. Field names verified in claude-meter's UsageResponse struct, src/models.rs lines 19-28.

Because four of the eight enforcement floats are not rendered on claude.ai/settings/usage. The most common silent cause is extra_usage.utilization (your metered overage hit its monthly cap) and the second-most is seven_day_oauth_apps (the OAuth-only bucket Claude Code traffic counts toward separately). The five_hour bar can read 40 percent, the three model bars can read 50 to 70 percent, and the next prompt still 429s because one of the four hidden floats is at 1.0.

Each bucket carries its own resets_at timestamp in the JSON. The five-hour window rolls continuously: when you sent your oldest message in the current window ages off, that message's cost drops out and the float decreases. resets_at on the five_hour field is the timestamp of that next age-off, not a cliff reset to zero. Weekly buckets behave the same way over seven days. ClaudeMeter prints the per-bucket reset clock as a local timestamp ('-> resets Sat May 16 18:42 (in 2h)') in src/format.rs lines 90-113.

Only the five_hour float did. Anthropic announced on May 7, 2026 that the 5-hour rate limit doubled for Pro, Max, Team, and seat-based Enterprise, and peak-hour throttling was removed for Pro and Max. The weekly caps (seven_day, seven_day_sonnet, seven_day_opus, seven_day_oauth_apps) were not changed. In practice many heavy users now hit the weekly walls earlier in the week instead of bumping into the 5-hour wall mid-flow.

ccusage reads ~/.claude/projects/<project>/<session>.jsonl on your machine and sums token counts against a price card. The plan limits Anthropic enforces live as the eight utilization floats on its server. The two are different data sources answering different questions. ccusage is good for 'what would these prompts have cost at API prices?' It cannot read the server quota because the server quota is not in your local logs. ClaudeMeter reads the server quota directly; running both is reasonable.

Install ClaudeMeter: brew install --cask m13v/tap/claude-meter, then load the browser extension from the GitHub releases page once. The extension polls /api/organizations/{org}/usage every 60 seconds using your existing claude.ai cookies (no password, no token, no manual cookie paste) and POSTs the parsed snapshot to a localhost-only bridge on 127.0.0.1:63762 (BRIDGE constant in extension/background.js line 2). The menu-bar app redraws the four visible bars plus the four hidden floats. You can also run /Applications/ClaudeMeter.app/Contents/MacOS/claude-meter --json for a one-shot JSON dump in a terminal or tmux status line.

Pro and Max only. The /api/organizations/{org_uuid}/usage endpoint is part of the claude.ai web account surface; it returns plan-utilization floats for subscription seats. API-key-only customers do not have a 'plan quota' in that sense, they have per-key rate limits that surface as 429 headers on the Anthropic API itself. ClaudeMeter is built for the subscription cohort because that is the cohort whose limits are invisible until they fire.

It means your last prompt would have crossed 1.0 on one of the eight floats. The CLI does not tell you which float because Anthropic's 429 response body does not disambiguate. The only ways to find out are: (a) open DevTools on claude.ai/settings/usage and check the network call to /usage to see which field is at 1.0, (b) curl http://127.0.0.1:63762/snapshots with ClaudeMeter running and inspect the parsed snapshot, or (c) run claude-meter --json directly. Knowing which float fired matters because the reset clock is different for each bucket.

ClaudeMeter is MIT-licensed open source; the entire data path is in /Users/<you>/claude-meter/src/. Two routes exist: the browser-extension route never touches your cookie jar (the extension calls the API with credentials: 'include' and posts the parsed snapshot to localhost), and the keychain route shells out to security find-generic-password to read Chrome Safe Storage on your own machine. No telemetry, no analytics, no network egress beyond claude.ai itself. The bridge listens on 127.0.0.1:63762 (loopback only), so nothing on your LAN can reach it.

The endpoint is undocumented and could change at any time. It has not changed in the year since claude.ai/settings/usage shipped. Because ClaudeMeter is a thin parser over the same JSON the Settings page renders, breakages tend to be field-shape changes that update with a one-line patch to src/models.rs. Anthropic has never blocked the endpoint behind a CSRF check or per-request token; the only auth is the same session cookie your browser already holds.