Claude Max usage tracking is seven counters, not one

Open claude.ai/settings/usage in a logged-in browser. The page reads three internal endpoints behind the scenes: /api/organizations/{org}/usage (returns seven utilization windows plus the extra_usage block), /api/organizations/{org}/overage_spend_limit (your monthly extra-usage cap and credits used), and /api/organizations/{org}/subscription_details (next charge date and payment method). To watch all three live, install ClaudeMeter; the macOS menu bar app and the browser extension both pull the same three endpoints with your existing claude.ai session cookies once a minute. No paste, no API key, no embedded sign-in.

From the JSON the endpoint returns: five_hour (the rolling 5-hour session window), seven_day (the overall weekly bucket), seven_day_sonnet (Sonnet-specific weekly bucket), seven_day_opus (Opus-specific weekly bucket), seven_day_oauth_apps (usage attributed to OAuth-connected third-party apps), seven_day_omelette (an internal-name bucket that some accounts have populated), and seven_day_cowork (collaboration features). They are independent. You can be at 12 percent on seven_day_opus and 91 percent on five_hour at the same moment. The rate limiter fires on whichever bucket hits 100 first, which is why watching only the first two collapses the picture.

Because Opus eats the Opus-specific weekly bucket faster than the same workload eats the Sonnet bucket, but most trackers only show one combined weekly bar. If you do a heavy Opus refactor on Monday morning, your seven_day_opus.utilization can be at 70 percent while seven_day.utilization is at 28 percent. The combined-weekly bar tells you you are fine. The Opus-specific bar tells you Opus is going to throttle this Wednesday. They are different questions.

It is the new April 2026 metered-billing block on the same /usage payload. The shape is { is_enabled, monthly_limit, used_credits, utilization, currency }. is_enabled tells you whether your account opted into pay-as-you-go; monthly_limit is the dollar cap you set; used_credits is how much you have spilled this month after running out of plan quota; utilization is used_credits / monthly_limit. ClaudeMeter parses this into an ExtraUsage Rust struct (src/models.rs lines 9 to 16) and surfaces it next to the seven utilization windows so you can see plan-quota exhaustion and dollars-spent in the same view.

No. Both read ~/.claude/projects/*.jsonl files written by the Claude Code CLI and count tokens locally. Local token counts have no relationship to the server-side utilization fields. ccusage is great for understanding which session burned which model and at what cost; it cannot tell you what percent of your Max plan's seven_day_opus bucket is gone, because that information only exists on Anthropic's side. ClaudeMeter and ccusage measure different things and complement each other.

Inconsistent. The same payload from the same endpoint returns some windows on the 0-1 scale and others on the 0-100 scale. ClaudeMeter normalizes with the clamp u <= 1 ? u * 100 : u (extension/background.js, pctFromWindow function, lines 58 to 63). If you build your own dashboard against this endpoint without that clamp, you will plot two scales on the same axis and the chart will look broken every time the format flips.

Once per 60 seconds. The browser extension uses chrome.alarms.create('refresh', { periodInMinutes: 1 }) (extension/background.js line 105). The macOS menu bar app refreshes on the same cadence by default and lets you change it from 30 seconds to 5 minutes in the popover. 60 seconds is the cadence that catches a slope change in five_hour without hammering Anthropic's endpoint.

No. claude.ai/settings/usage renders a 5-hour bar, a weekly bar, and the per-model breakdowns when expanded. seven_day_oauth_apps, seven_day_omelette, and seven_day_cowork are in the JSON response but are not surfaced as separate bars on the page. You see them by opening DevTools, copying the response of /api/organizations/{org}/usage, and reading the field list. ClaudeMeter shows whichever ones come back populated for your account so a non-zero bucket cannot hide.

No. It is internal and undocumented. It powers claude.ai/settings/usage, which means Anthropic depends on it being correct, but they can rename fields or remove buckets in any release. ClaudeMeter deserializes into an explicit Rust struct (src/models.rs UsageResponse), so a schema change surfaces as a parse error rather than a silent zero. The seven field names listed above were stable through 2026-04-29.

There is nothing for it to meter on a free account. The endpoint returns a payload with the windows present but utilization fields populated only when the plan has quota to consume. If you are on Pro you see five_hour and seven_day plus the per-model windows. If you are on Max you see all seven plus extra_usage when the metered-billing block is enabled. Free accounts do not have plan quota, so the bars stay at zero or null.