[BUG] Per-prompt cap on claude-opus-4-7 silently dropped ~2x around Apr 27, 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?

The per-prompt size cap on claude-opus-4-7 appears to have silently dropped
~2x between Apr 26 and Apr 27, 2026. Sessions that previously held 600-750k
cached tokens with zero rejection are now getting rejected at 360-440k. Same
model, same account, no behavior change on my end.

Visible symptom: the red "Context limit reached" banner appears in Claude
Code at ~36-44% of the displayed 1M context_window. CC's context display
shows plenty of headroom; the API rejects anyway.

What Should Happen?

Either:

• The per-prompt cap should remain where it was — Opus 4.7 used to clear

600-750k tokens cleanly, and that's the implicit contract of the "1M
context window" framing.

• If the cap was intentionally lowered, the new limit should be documented

as a separate number from context_window_size and exposed in the API
response so clients can show users where the operational ceiling actually
is. Today the user sees context_window.used_percentage: 38 right up until
the wedge.

Error Messages/Logs

API response (synthetic assistant message in the session JSONL with
  isApiErrorMessage: true):                                                                                             
  
    "Prompt is too long"                                                                                                
    error: "invalid_request"                                                 
    model: <synthetic>                                                                                                  
    stop_reason: "stop_sequence"                                                                                        
                                         
  CC UI: red "Context limit reached" banner.                                                                            
                                                                             
 Evidence across 7 consecutive sessions, same account, same model                                                      
  (claude-opus-4-7), CC v2.1.107 → v2.1.121:                                          
                                                                                                                        
  | Session ID | Date   | Max cache_read_input_tokens | "Prompt is too long" errors |                                   
  |------------|--------|------------------------------|------------------------------|                                 
  | f4955755   | Apr 23 |  87k                         |  0                           |                                 
  | 0ec73e4b   | Apr 23 | 689k                         |  0                           |                                 
  | 6173cc94   | Apr 24 | 692k                         |  0                           |                                 
  | 346c895b   | Apr 25 | 636k                         |  0                           |                                 
  | 6a7ed1d9   | Apr 26 | 752k                         |  0                           |                                 
  | e696482b   | Apr 27 | 441k                         | 11                           |                                 
  | 1d546a39   | Apr 28 | 363k                         |  1                           |                                 
                                                                                                                        
  JSONL transcripts available on request. Pulled from                                                                   
  ~/.claude/projects/<slug>/<session-id>.jsonl using:                                                                   
                                                                                                                        
    grep -oE '"cache_read_input_tokens":[0-9]+' <file> | sort -t: -k2 -rn | head                                        
    grep -c '"isApiErrorMessage":true' <file>                                                                           
                                                                                                                        
  Last rejected call this morning: cache_read_input_tokens=363036, the prompt                                           
  was rejected with "Prompt is too long".

Steps to Reproduce

This is a state condition, not a deterministic code path. Best approximation:

1. Open a Claude Code session on claude-opus-4-7.
2. Use a project with a substantial CLAUDE.md (~17k tokens in my case) and

normal auto-memory loading.

1. Work normally for a session — file reads, bash greps, back-and-forth.

Avoid manual /compact.

1. As cache_read_input_tokens climbs past ~360-400k, the next API call is

rejected with "Prompt is too long" even though
context_window.used_percentage shows ~36-44%.

1. Verify in the session JSONL:

grep '"isApiErrorMessage":true' <session.jsonl>

and cross-reference the cache_read_input_tokens on the immediately
preceding successful turn.

The exact rejection point varies (363k, 382k, 441k observed across sessions)
so a deterministic repro may not be possible — but the cross-session
ceiling drop is unambiguous: pre-Apr 27 sessions cleared 600-750k cleanly,
post-Apr 27 sessions reject at 360-440k. Manual /compact recovers
immediately.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.1.121

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_