Opus 4.7 regression vs 4.6: ~2x token burn, mid-session stalls, no quality gain

Summary

Paying Claude Code user reporting concrete regressions on Opus 4.7 (1M context) vs Opus 4.6 in daily use.

Observed regressions

1. No perceptible capability gain over 4.6. For daily coding/agent tasks (TypeScript, Shopify app work, CI debugging) 4.7 produces output of similar or lower quality than 4.6.
2. ~2× token consumption for equivalent tasks. Usage cost has roughly doubled without a matching quality improvement.
3. Mid-session stalls. 4.7 gets stuck mid-session (tool loop / no forward progress) noticeably more often than 4.6. Session restart required to recover.
4. UI/harness regression. The recent Claude Code interface changes coincide with models behaving worse — shorter reasoning, more truncation, more confident-but-wrong answers. Whatever changed in the harness/system prompt around the UI refresh appears to have hurt model behavior.

Request

• Investigate the 4.7 regression vs 4.6.
• Keep 4.6 selectable as a stable fallback while 4.7 is being addressed.

Environment

• Claude Code CLI (Windows)
• Model: claude-opus-4-7 (1M context)
• Account email: bsdaom778@gmail.com