Model quality degradation since Jan 8 - decreased understanding, verbose output, faster but worse

Summary

Noticed significant model quality degradation starting January 8, 2026 (UTC+8), ongoing for 3 days now.

Timeline

• First noticed: 2026-01-08 17:00-19:30 (Singapore time)
• Jan 9: Issues continued
• Jan 10: Problems became more severe, especially early morning

Symptoms

| Symptom | Description |
|---------|-------------|
| Understanding decreased | Model misinterprets instructions more frequently |
| Verbose output | More filler content, less substance |
| Faster output | Noticeably faster generation, but shallower reasoning |
| Higher token burn | More tokens consumed for same tasks |

Behavioral Changes

The model feels like it switched from "think then speak" to "speak while thinking" - or not thinking at all.

• Less <pause> moments (metaphorically)
• More System 1 (fast/intuitive) responses
• Less System 2 (slow/deliberate) reasoning

Possibly Related Issues

These symptoms appeared around the same time as:

• #14941 - Compact too aggressive
• #16509 - Interrupt not working during tool execution
• #14176 - Compact loses TodoList state

Questions

1. Was there a model change or parameter adjustment around Jan 8?
2. Is this an A/B test or full rollout?
3. Are others experiencing similar degradation?

Environment

• Heavy daily usage (12+ hours)
• Complex multi-file coding tasks
• Using Claude Code CLI

---

Reporting this because 3 days of degradation suggests this is intentional or unnoticed, not a temporary glitch. Happy to provide more specific examples if helpful.