A programmatic Claude usage tracker before June 15, written in one CLI command against an endpoint Anthropic does not document.

Anthropic retires the original Claude Sonnet 4 (claude-sonnet-4-20250514) and Claude Opus 4 (claude-opus-4-20250514) on June 15, 2026. After that date, requests to those model IDs return errors. The replacements are Sonnet 4.5 / 4.6 and Opus 4.5 / 4.7. The official deprecation page is at platform.claude.com/docs/en/about-claude/model-deprecations. If you have any pipeline still pinned to the 20250514 model strings, this is the calendar date it stops working.

Two reasons. First, the new models are weighted differently against the rolling 5-hour and weekly windows; an Opus 4.7 turn does not consume the same fraction of plan as an Opus 4 turn did. Anything you measured pre-migration is a stale baseline. Second, anyone running model-pinned scripts on a Pro or Max subscription needs a way to see the new burn rate from a script, not just from a settings page. That is the case the consumer plan never had a real answer for, and where the OAuth endpoint comes in.

Yes, two of them: the Usage & Cost Admin API at /v1/organizations/usage_report/messages and the Enterprise Analytics API. Both authenticate with an Admin API key (sk-ant-admin...), both report per-message token consumption and dollar cost for API-key traffic. Neither one exposes the consumer Pro / Max plan's 5-hour and 7-day windows. Those are a separate enforcement surface. If your usage runs through Claude Code, claude.ai chat, or anything else billed against a Pro / Max subscription rather than an API key, the official Usage & Cost API will not see your traffic.

On two undocumented endpoints. claude.ai/settings/usage renders from GET /api/organizations/{org_uuid}/usage with your browser session cookie. The same shape is reachable on api.anthropic.com via GET /api/oauth/usage with the OAuth Bearer token Claude Code already stashed in your macOS Keychain under service "Claude Code-credentials". Both routes return the same eight-float UsageResponse: five_hour, seven_day, seven_day_sonnet, seven_day_opus, seven_day_oauth_apps, seven_day_omelette, seven_day_cowork, plus an extra_usage block for the metered overage balance.

A JSON array of UsageSnapshot objects. Each snapshot has org_uuid, browser, account_email, fetched_at, and the parsed usage / overage / subscription blocks. The shape is defined in src/models.rs lines 60-73. A typical day's array is about 1.5 KB, easy to pipe into Slack, a Starship segment, a tmux status line, an Alertmanager rule, or whatever else you script against. There is no API key to manage; the tool reads Claude Code's existing OAuth credentials and, if those are missing or expired, falls back to your already-logged-in browser cookies.

It is the same endpoint the running Claude Code CLI calls every time it refreshes its own usage cache. The Bearer token is the one Claude Code minted for you when you logged in (scope user:profile is enough). claude-meter does not refresh the token (Claude Code does that automatically with its refresh token); it just reads what is already in the keychain. The cadence is one request per minute, well under what an open Claude Code session generates. The endpoint is undocumented, which means Anthropic can rename a field and break the parse; not that they prohibit reading it.

Yes. Read the keychain entry, pull the accessToken field, and curl the endpoint yourself: security find-generic-password -s 'Claude Code-credentials' -w | jq -r .claudeAiOauth.accessToken | xargs -I {} curl -s https://api.anthropic.com/api/oauth/usage -H 'Authorization: Bearer {}' -H 'anthropic-version: 2023-06-01'. You get the same JSON. claude-meter just wraps that with token-expiry handling, multi-account dedupe, an optional menu bar surface, and a polished --json shape.

Today, claude-meter is macOS only because the keychain reader is /usr/bin/security and the cookie fallback understands Chromium's Safe Storage on macOS. The OAuth endpoint itself is not platform-specific, so the same curl pattern works on Linux or Windows if you can locate Claude Code's stored credentials on your platform (Claude Code uses the OS keychain on each: libsecret on Linux, Credential Manager on Windows). The lugia19/Claude-Usage-Extension Chrome extension is the cross-platform option that reads the cookie path instead.

Yes. They answer different questions. ccusage walks ~/.claude/projects/*.jsonl and totals the tokens your local Claude Code CLI sent to Anthropic. claude-meter reads the server's plan-side utilization. ccusage tells you what your machine spent in tokens; claude-meter tells you what fraction of plan ceiling Anthropic counted that against. Local-token sums and server utilization can drift by 30-40 points and both still be correct readings of what they measure. Run both. They poll different sources.

No. The model deprecation removes the old model identifiers from the inference API; it does not touch the usage endpoint or the OAuth scope. claude-meter keeps reading the same JSON; the only thing that changes is the Sonnet / Opus weekly buckets start filling at the 4.5 / 4.6 / 4.7 burn rate instead of the 4.0 burn rate. The breakage to plan for is at the inference layer of any pinned-model script you run, not at the metering layer.