Claude users didn’t imagine it. Their AI companion turned forgetful. Repetitive. Dull. Complaints flooded forums through March and April. Then Anthropic stepped up. The company pinned the blame on three specific changes gone wrong in Claude Code, the Agent SDK, and Cowork. The API? Untouched. On Thursday, April 23, Anthropic dropped a postmortem. Engineers detailed the mess , insisting no deliberate dumbing down occurred. Just missteps. Cost-saving tweaks that backfired spectacularly. First culprit: March 4. Anthropic dialed Claude Code’s default reasoning effort from high to medium. Why? Latency. High effort meant longer thinking, frozen UIs. Users wanted smarts over speed, though. “This was the wrong tradeoff,” Anthropic admitted. They flipped it back April 7. Now v2.1.118 defaults to “xhigh” on Sonnet 4.6. Lighter inference load? Sure. But at quality’s expense. Next. March 26. A cache tweak. Claude caches input tokens for an hour—faster, cheaper calls.…