[Bug] Anthropic API: Extreme latency on Claude Code requests (15s+ to first token)

Bug Description
Claude Code extremely slow response times on Claude Max plan (since March 25)
Plan: Claude Max
Claude Code version: 2.1.85
Model: Opus 4.6 (also tested Sonnet 4.6 — same issue)
OS: macOS (Ghostty terminal)
Issue
Since approximately March 25, Claude Code has become unusable due to extremely slow response times. Simple queries that should take seconds are taking 1–5+ minutes. The Claude Desktop app works noticeably faster with the same account.
Example
A simple query like "find the path for MVContactCard component" — which only requires a single Glob search — takes 5 minutes 25 seconds to complete. The Glob search itself finishes instantly; the delay is entirely in the API response.
Debugging findings
I enabled debug mode (CLAUDE_CODE_DEBUG=1) and captured detailed logs. Here's what I found:
Network is not the issue:
curl to api.anthropic.com:
TCP connect: 0.048s
TTFB: 0.105s
Total: 0.106s
API response timeline for a simple file search query:
08:05:55.363 — API request sent (repl_main_thread)
08:06:10.346 — Stream started, first chunk received (15s to first token)
08:06:10.488 — Glob tool executed (instant)
— NO second API request logged in debug output
— Response finally completed after 5m 25s total
The debug log shows no 529 errors, no retries, and no second API request after the tool use. The API simply takes minutes to generate a response that is two sentences long.
529 Overloaded errors also observed:
On a separate attempt, I received an explicit 529 error:
API Error: 529 {"type":"error","error":{"type":"overloaded_error","message":"Overloaded"}}
Claude Code retried for 8 minutes 20 seconds before giving up.
Context size is minimal:

Context usage: 4%
Usage quota: 1% (fresh 5-hour window)
Session was freshly cleared with /clear before testing

Desktop app is faster:
The same account on the Claude Desktop app (Cowork) responds in reasonable time. The issue appears specific to Claude Code or its API routing.
Not model-specific:
Tested with both Opus 4.6 and Sonnet 4.6 — both exhibit the same slow behavior.
Project configuration
My project does load a significant amount of context (38 CLAUDE.md/rules files, 87 skills from 8 plugins, 10+ MCP servers), but this same configuration worked fine before March 25. The project context occupies only 4% of the 1M context window.
Environment

Claude Code v2.1.85
Claude Max plan
macOS, Ghostty terminal
Project: ~7,900 tracked files

Request
Could you please investigate why Claude Code API responses are taking minutes for trivial queries on the Max plan? The status page shows all green, but the experience has been severely degraded for multiple days. This appears related to the issues reported in GitHub issues #38335, #39767, #39745, and #39747.

Environment Info

• Platform: darwin
• Terminal: ghostty
• Version: 2.1.86
• Feedback ID: 223885e5-4340-43ad-94c6-c73a65be61f0

Errors

[]