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. Engineers aimed to dump idle output tokens, those “thinking sessions,” after an hour. Cut resumption costs. A bug hit instead. Cleared thinking every turn. Claude forgot. Repeated itself endlessly. Fixed April 10 for Sonnet 4.6 and Opus 4.6.
Last straw. April 16. System prompt rewrite for brevity. “Length limits: keep text between tool calls to ≤25 words. Keep final responses to ≤100 words unless the task requires more detail.” Internal tests passed. But paired with Opus 4.7 release, ablation tests exposed a 3% performance plunge across Opus 4.6 and 4.7. Reverted April 20.
These hit staggered. Different models, traffic slices. Mimicked creeping decline. Reddit ranted. X erupted. The Register captured the fury earlier. Head of growth Amol Avasare addressed an unannounced A/B test on X, vowing direct talk over social sleuthing.
Anthropic acted. Reset all usage limits. Promised rigorous evals, Code Review upgrades, a @ClaudeDevs X handle for transparency. “This isn’t the experience users should expect from Claude Code,” they wrote.
Users tested fixes fast. Some cheered restoration. Others grumbled. “The fixed Claude is even worse,” one X post griped April 24. Another: Claude reset limits, but Opus 4.7 still refuses prompts, jailbreak-heavy safeguards intact. Timing? Ironic. OpenAI’s GPT-5.5 hype swirled same day.
Hacker News dissected the postmortem. Threads questioned intent—cost cuts disguised as bugs? Boris from Claude Code team jumped in, explaining context elision experiments. Thinking traces cleared poorly. Users gaslit pre-fix, per screenshots.
One HN thread tallied gaslighting examples post-April 10 “fixes.” Anthropic fumbled communication, critics said. Yet transparency won nods. Most firms ghost.
Broader context. Claude Code rides Opus 4.6/4.7, Sonnet 4.6 waves. Not the API’s raw power. Harness changes—effort levels, prompts, caches—shape output. Docs confirm: effort adjustable. But defaults matter.
Industry watches. VentureBeat flagged prompt injection in Claude Code Security Review April 23. Researcher exploited GitHub PR titles, leaked API keys. Anthropic paid $100 bounty, CVSS 9.4 critical. Patched quietly. System cards warn: not injection-hardened.
Ars Technica noted Pro plan A/B tests removing Code access. Users throttled. Trust eroded. And Mythos? Anthropic’s vuln-hunting beast found Firefox 150’s 271 zero-days, per Mozilla. But its own breach—basic access slip—humiliated, The Verge reported April 24.
Lessons stack. AI firms chase scale, safety. Tradeoffs bite. Users pay. Anthropic rebounded with resets, candor. Will it stick? Engineers bet on evals, communication. Bugs fixed. Confidence? Rebuilding.
Short term: Update to v2.1.118+. Test hard tasks. Long term: Watch @ClaudeDevs. Demand proof.
WebProNews is an iEntry Publication