[BUG]

### Preflight Checklist

- [x] I have searched [existing issues](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20state%3Aopen%20label%3Abug) 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?

Follow-up on conversation ID 215475713330029.

I extracted per-call cache accounting from my local Cowork session logs
(session "LiveProfit project handoff", Max 20x). Timestamps below are
UTC; I am in Europe/Zurich (UTC+2).

The cache was working normally, then broke mid-session with no change
on my side:

01:19:11 opus-4-8 cache_creation= 0 cache_read= 894,582
01:20:04 opus-4-8 cache_creation=722,972 cache_read= 43,876

From 01:20:31 onward, cache_read stayed frozen at exactly 43,876 for
~100 consecutive calls, while cache_creation rewrote ~750,000 tokens on
every single call — for responses of 17-20 output tokens.

Three ordinary explanations are ruled out by the same log:
- Model switching: none. Every call in this window is opus-4-8.
- Cache TTL expiry: calls are 7-8 seconds apart.
- Oversized context: 894,582 tokens were being read from cache one
minute earlier without issue.

Cost per call went from 388 tokens of cache_creation at 01:19 to
~750,000 at 01:24 — a factor of roughly 1,900.

The same failure recurred at 08:54 UTC on claude-fable-5, broken from
the first call, with cache_read frozen at 44,377.

Converted to local time, these two windows are 03:15-03:35 and
10:54-11:02 Europe/Zurich. My usage dashboard screenshots are timestamped
03:34 (session limit reached) and 11:20 (All models 100%) — attached.

Daily context: cache_creation was 2.0-2.6% of total tokens on Aug 27,
2.5% on Aug 28, 2.1% on Aug 29, and 76.4% (opus-4-8) / 87.9% (fable-5)
on Aug 30. Across three months of logs this session never exceeds 10%.

Result: my entire Max 20x weekly quota was consumed, no deliverable was
produced, and I was locked out for five days.

This matches the cache invalidation failure mode described in Anthropic's
April 2026 postmortem covering Claude Code, the Agent SDK and Cowork,
after which usage limits were reset for all subscribers. I am requesting
the same remedy: a usage limit reset or an equivalent account credit.

Raw session logs available on request.

### What Should Happen?

I dont know

### Error Messages/Logs

```shell

```

### Steps to Reproduce

Follow-up on conversation ID 215475713330029.

I extracted per-call cache accounting from my local Cowork session logs
(session "LiveProfit project handoff", Max 20x). Timestamps below are
UTC; I am in Europe/Zurich (UTC+2).

The cache was working normally, then broke mid-session with no change
on my side:

01:19:11 opus-4-8 cache_creation= 0 cache_read= 894,582
01:20:04 opus-4-8 cache_creation=722,972 cache_read= 43,876

From 01:20:31 onward, cache_read stayed frozen at exactly 43,876 for
~100 consecutive calls, while cache_creation rewrote ~750,000 tokens on
every single call — for responses of 17-20 output tokens.

Three ordinary explanations are ruled out by the same log:
- Model switching: none. Every call in this window is opus-4-8.
- Cache TTL expiry: calls are 7-8 seconds apart.
- Oversized context: 894,582 tokens were being read from cache one
minute earlier without issue.

Cost per call went from 388 tokens of cache_creation at 01:19 to
~750,000 at 01:24 — a factor of roughly 1,900.

The same failure recurred at 08:54 UTC on claude-fable-5, broken from
the first call, with cache_read frozen at 44,377.

Converted to local time, these two windows are 03:15-03:35 and
10:54-11:02 Europe/Zurich. My usage dashboard screenshots are timestamped
03:34 (session limit reached) and 11:20 (All models 100%) — attached.

Daily context: cache_creation was 2.0-2.6% of total tokens on Aug 27,
2.5% on Aug 28, 2.1% on Aug 29, and 76.4% (opus-4-8) / 87.9% (fable-5)
on Aug 30. Across three months of logs this session never exceeds 10%.

Result: my entire Max 20x weekly quota was consumed, no deliverable was
produced, and I was locked out for five days.

This matches the cache invalidation failure mode described in Anthropic's
April 2026 postmortem covering Claude Code, the Agent SDK and Cowork,
after which usage limits were reset for all subscribers. I am requesting
the same remedy: a usage limit reset or an equivalent account credit.

Raw session logs available on request.

### Claude Model

None

### Is this a regression?

Yes, this worked in a previous version

### Last Working Version

_No response_

### Claude Code Version

opus and fable

### Platform

Anthropic API

### Operating System

macOS

### Terminal/Shell

Terminal.app (macOS)

### Additional Information

_No response_

Start with conversation ID 215475713330029 and the reported Cowork session logs, comparing the cache_read and cache_creation values across the two failure windows. Review the attached usage-dashboard screenshots and the referenced cache-invalidation postmortem; done would be a confirmed cause and a decided remedy for the consumed quota, such as a limit reset or equivalent account credit.

索引モデルが issue の本文から書いたものです。