Anthropic gave you credits. The shortage and the cache TTL drop turned them into funny money.

It is shorthand for a balance that looks real on a settings page but cannot reliably be spent in the moment a user wants to spend it. In the Anthropic 2026 case, the make-good credits (a one-month subscription refund, a $200 extra-usage promo, monthly metered allowances) are denominated in real dollars. They sit in /api/organizations/{org_uuid}/overage_spend_limit. But three things conspire against you actually using them: the rolling-window throttle stops you before you can spend down the credit, the prompt cache change inflates the per-prompt burn so the credit drains faster than the work you got out of it, and the server can flip an out_of_credits flag that disables the line entirely. The dollars are technically yours; the spending behaviour is not.

In the OverageResponse JSON returned by GET https://claude.ai/api/organizations/{org_uuid}/overage_spend_limit. The full struct is seven fields: is_enabled, monthly_credit_limit, currency, used_credits, disabled_reason, disabled_until, out_of_credits. ClaudeMeter decodes that struct in src/models.rs lines 30-40 and renders the boolean as the string BLOCKED in src/format.rs line 26. Three of those fields, out_of_credits, disabled_reason, and disabled_until, are not visible anywhere a local-log tool reaches.

The default Claude Code prompt cache TTL dropped from 1 hour to 5 minutes around April 2-3, 2026. The rollout was staggered, so different users saw the change on different days; some retained 1-hour caching well past the cutover. One Reddit user documented the impact precisely: 39 cache busts per day at $6.28 before April 2, and 199 cache busts per day at $15.54 after, a $277.80 monthly delta on identical work. xda-developers and The Register covered the change after the fact. The reason it matters here is that those extra cache rebuilds get charged at the cache-write rate against the same credit balance Anthropic gave you to make up for the shortage.

Anthropic tightened consumer-plan rolling windows on weekday peak hours (8am to 2pm Eastern) starting late March 2026. The product lead acknowledged users were 'hitting usage limits in Claude Code way faster than expected' on March 31. When the 5-hour or 7-day window trips, the Claude Code client falls through to extra-usage billing, which spends down your make-good credit instead of returning a 429. So 'I bought you a month' becomes 'I bought you a few peak-hour blocks of throttled inference', and you have no way to see the substitution happening unless you read the right server field.

Because they read the local JSONL transcript on disk and sum input/output tokens. /overage_spend_limit is a different endpoint on a different host (claude.ai, not platform.claude.com). It returns dollar amounts in cents, plus the three diagnostic fields above. There is no token field on that endpoint, and there is no overage field in the local transcript. The two data sources do not overlap; a tool that walks JSONL has no way to find out the credit line is BLOCKED until you fail a request and read the error body.

src/format.rs lines 24-40. The CLI prints a line like '$12.40 / $200.00 (6%) BLOCKED until Sat May 10' when out_of_credits is true and disabled_until is set. The menu bar reads the same struct and surfaces the BLOCKED string under the extra-usage row. If disabled_reason is present, the binary appends it. The point is that a single field flip on the server changes one boolean in OverageResponse and the UI reflects it on the next 60-second poll.

No. It is the same endpoint the claude.ai/settings/usage page renders from, and Anthropic does not publish its shape. ClaudeMeter is open source under MIT and the struct is at /Users/matthewdi/claude-meter/src/models.rs (or github.com/m13v/claude-meter/blob/main/src/models.rs). The JSON parse uses serde with strict typing, so if Anthropic renames a field, the next poll fails loudly with the parse error rather than silently misreporting.

There is no automatic refund mechanism. The official help center page on requesting a refund covers paid plan refunds, not extra-usage credit reversals. Reading the disabled_reason field on the day it flips is the closest thing to evidence; ClaudeMeter logs the snapshot to a localhost bridge, so you can grep your own historic snapshots if you ever need to argue the case with support. We are not a billing claim service; we just expose the field.

No. The Console API on platform.claude.com has its own Usage and Cost endpoint, its own anthropic-ratelimit-* response headers, and its own credit ledger. The credits you buy on the API console are separate from the make-good credits Anthropic dropped on consumer plans. The shortage and the cache TTL change have hit Claude Code (which runs on consumer auth in many flows) and claude.ai chat. If you only consume via the public API, you are seeing a different surface and this page is not for you.

Three things. One: install ClaudeMeter and the browser extension; the menu bar shows your extra-usage dollars and any BLOCKED suffix without you opening the settings page. Two: keep the snapshots the bridge writes to localhost; they are JSON timestamps you can grep later. Three: when you hit a hard wall, check whether the wall is a 5-hour rolling window (the bucket on /usage), a weekly window (the seven_day bucket), or extra-usage being disabled (out_of_credits true on /overage_spend_limit). They look the same in Claude Code's error message; they have different recoveries.