Did Claude really get dumber again?

A detailed investigation into why Claude models (especially Opus 4.6/4.7) appear to be getting dumber over time. The author identifies multiple contributing factors: poor harness engineering in Claude Code (Opus performs 15% worse in Claude Code vs Cursor), a tokenizer change in Opus 4.7 that increases token usage by up to 1.47x causing context bloat, the rollout of 1 million token context windows as default (which Anthropic's own postmortem confirmed degrades quality), and the redaction of thinking blocks starting March 5th that reached 100% by March 12th. AMD's AI director published quantitative analysis of 17,000+ thinking blocks showing 73% reduction in thinking depth, 80x increase in API requests, and a read-to-edit ratio dropping from 6:1 to 2:1. The author concludes Anthropic's engineering culture is fundamentally broken and recommends switching to alternatives like Codex.