[BUG] Anomalous usage meter burn rate following Jul 6 claude.ai incident — 5% per short message on Max 5x, all models affected, Fin support response mathematically impossible

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Following the claude.ai elevated errors incident on Jul 6, 2026 (19:18–20:05 UTC), my usage meter began reporting impossible consumption rates. One 3-sentence message to Sonnet 5 Low/Thinking OFF consumed 5% of a fresh Max 5x session. Switching to Sonnet 4.6 (no new tokenizer) produced identical rates. This is not a Claude Code CLI issue — I'm filing here as it is Anthropic's only public issue tracker for claude.ai consumer billing bugs.

What Should Happen?

A 3-sentence message to Sonnet 5 Low/Thinking OFF should consume well under 1% of a Max 5x session, consistent with 7 days of normal usage on this account with this model since June 30.

Error Messages/Logs

No error messages. Usage meter silently reports impossible consumption. Support Conversation ID: 215474981789339

Steps to Reproduce

1. Account experienced the Jul 6 19:18 UTC claude.ai incident (full lockout, zero messages sent)
2. Incident resolved at 20:05 UTC — session showed 93% consumed
3. Session reset — sent one 3-sentence message, Sonnet 5 Low/Thinking OFF — meter showed 5% on Max 5x
4. Switched to Sonnet 4.6, same settings — identical burn rate on Max 5x
5. After next reset, first message = 5%, second message = 11% on Max 5x

Claude Model

Sonnet (default)

Is this a regression?

Yes, this worked in a previous version

Last Working Version

claude.ai web, July 5 2026

Claude Code Version

N/A

Platform

Other

Operating System

Windows

Terminal/Shell

Other

Additional Information

Mathematical proof tokenizer explanation is impossible: 5% per message on Max 5x = ~25% per message on Pro = over 100% per message on Free. Free users are actively sending messages. Therefore this cannot be real token cost. Related issues: #41930 (usage drain since March), #57594 (Fin checks wrong billing record). Fin AI Agent response was mathematically impossible and disproven by controlled test (switching models produced identical rates).

<img width="1673" height="995" alt="Image" src="https://github.com/user-attachments/assets/74fd0ee5-cd1a-48dc-a02b-4b9dd9e652d7" />