[BUG] Opus 4.6 extremely slow inference — 10+ minutes for 2k tokens on Max plan

Bug Description

Opus 4.6 inference is extremely slow — 9-10+ minutes for ~2k tokens output. This happens consistently, not occasionally. Simple tasks that should take seconds hang on "Inferring…" for 5-10 minutes.

Example from today: after a Bash command completed, Claude spent 10m 28s "Inferring" to generate a ~2k token response.

This matches the symptoms reported in #34623, #25059, and #24585.

Environment Info

• Platform: macOS (Darwin 24.5.0)
• Version: 2.1.81
• Model: Claude Opus 4.6 (1M context)
• Plan: Claude Max
• MCP servers connected: 12 (Attio, Google Calendar, Gmail, Notion, Grain, Playwright, Context7, Figma, Webflow, Google Workspace, Open Brain x2)
• Shell: zsh

Steps to Reproduce

1. Start a conversation on Opus 4.6 with Max plan
2. Run any multi-step task (file edits + bash commands)
3. After a tool call completes, observe "Inferring…" for 5-10+ minutes
4. Output is only ~2k tokens

Expected Behavior
Inference should complete in seconds for 2k tokens, not 10 minutes.

Additional Context

• Tried disabling MCPs — no improvement
• Anthropic status page shows no incidents
• Sonnet does not exhibit this behavior
• Problem is hit-or-miss per #34623 but trending worse over time