[BUG] Opus 4.7 cache hit rate collapse after May 27 incident — Messages 1.1k→88.9k in 9 minutes, $630/session

Resolved 💬 2 comments Opened May 28, 2026 by garaz1983-rgb Closed Jul 2, 2026

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?

After the May 27 Opus 4.7 "elevated errors" incident (status.claude.com),
cache hit rate collapsed from ~95% to ~50%, causing extreme token consumption.

  • After running /compact, Messages started at 1.1k tokens
  • 9 minutes later it reached 88.9k tokens — with only one short user input

("finish the previous edit")

  • The project is a lightweight mobile app (small Kotlin codebase)
  • Single session cost: $630.42 API-equivalent (377 calls)
  • Full day (May 28): $1,142.35 (4,226 calls) on Max 20x ($200/month) plan

Cache hit rate comparison (CodeBurn dashboard):

  • Opus 4.7, 7-day average: 95.7%
  • Opus 4.7, May 29 only: 50.9% ← broken
  • Opus 4.6, same period: 89.9% ← normal

Autocompact buffer stayed fixed at 33.0k regardless of Messages growth.
A fresh session on the same project shows Messages at 13 tokens (normal baseline).

Environment: Windows 11, VS Code extension, claude-opus-4-7, Max 20x plan

What Should Happen?

Messages should grow proportionally to actual tool calls and user input.
Cache hit rate should remain ~95% as observed before the May 27 incident.
A short instruction on a small project should not consume 88k tokens in 9 minutes.

Error Messages/Logs

No error messages — the issue is silent.
  Token consumption explodes without any visible warnings.
  Only observable through Context usage panel or external tools like CodeBurn.

Steps to Reproduce

  1. Use Claude Code with Opus 4.7 in VS Code extension (Max 20x plan)
  2. Open any small project
  3. Run /compact to reset Messages to ~1k
  4. Give a short follow-up instruction (e.g. "finish the previous edit")
  5. Open Context usage panel repeatedly over the next 10 minutes
  6. Observe Messages growing ~10k/minute without proportional user input

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

Worked normally until May 27, 2026

Claude Code Version

2.1.80 (Claude Code)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

VS Code integrated terminal

Additional Information

The May 27 "elevated errors on Claude Opus 4.7" incident (status.claude.com)
was marked resolved, but cache behavior has not recovered.

Observed across multiple projects — not limited to one session or project.
Messages reached 385k+ in other projects on the same day.

Screenshots attached: Context usage timeline (3:17→3:26),
CodeBurn dashboard (today + 7-day), fresh session baseline.

<img width="688" height="620" alt="Image" src="https://github.com/user-attachments/assets/09ee970b-5a30-43a0-94af-aa0a678f582b" />
<img width="485" height="507" alt="Image" src="https://github.com/user-attachments/assets/1951c3a2-e857-43df-bc59-72ccde96a9ae" />
<img width="502" height="525" alt="Image" src="https://github.com/user-attachments/assets/ec789aed-695c-4e7c-ac41-a9a53308d7ee" />
<img width="480" height="476" alt="Image" src="https://github.com/user-attachments/assets/294806e9-5b17-4352-a579-7a2ed85f861c" />
<img width="450" height="469" alt="Image" src="https://github.com/user-attachments/assets/2d08681a-7889-4340-94ea-b5534fecd8fa" />
<img width="444" height="474" alt="Image" src="https://github.com/user-attachments/assets/a6607f52-87cd-4f7c-9f93-d5f9b13ae7c4" />
<img width="1860" height="1031" alt="Image" src="https://github.com/user-attachments/assets/08d6905e-d318-4219-a808-09e218e3e108" />
<img width="1632" height="1002" alt="Image" src="https://github.com/user-attachments/assets/574f97ef-b8b9-4b12-a03d-4689c3d644e4" />
<img width="720" height="715" alt="Image" src="https://github.com/user-attachments/assets/345804da-568e-40a0-af75-e4b01469d1c5" />

View original on GitHub ↗

This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗