If you've been wondering why Claude started acting like it forgot how to code mid-session, you're not losing your mind—Anthropic just confirmed it. The company released a detailed post-mortem on May 13, 2026, admitting that three separate infrastructure changes quietly degraded performance for weeks before anyone outside the loop caught on.

Mistake #1: Trading Depth for Speed

The first culprit? A stealthy adjustment to Reasoning Effort. Anthropic lowered it from High to Medium behind the scenes, ostensibly to reduce screen latency and make responses feel snappier. The trade-off was brutal—Claude stopped doing the deep analysis it's known for and started taking shortcuts instead. If you were relying on Claude to debug critical systems during this window, you might have been handed confidently wrong answers without any warning that the model's behavior had fundamentally changed.

Mistake #2: The Memory Leak That Erased Conversations

A bug in the caching layer was supposed to fire once when a session expired. Instead, it triggered every single turn. The result? Claude lost the entire conversation context each time you sent a follow-up message. Users reported having to re-explain requirements repeatedly, watching the model loop back to solutions they'd already rejected. This wasn't a quirk—it was a full-on memory wipe happening silently in production.

Mistake #3: Forcing Concise Output Stripped Necessary Reasoning

Anthropic added a system prompt designed to reduce verbosity and cut token costs. The unintended consequence was Claude cutting out essential reasoning steps and critical details, especially in code generation tasks. Quality dropped measurably because the model was optimizing for shortness instead of correctness.

Everything's Fixed—But the Transparency Is What Matters

All three issues were patched as of April 20, 2026 (patch v2.1.116). Reasoning Effort is back to High, with Opus 4.7 specifically configured to 'xhigh' as compensation. Anthropic also issued Usage Limit refunds starting April 23. But here's what actually matters: they told the truth about it. Full transparency, no spin.

Key Takeaways

  • Stealth changes to AI behavior without user notification erode trust fast
  • Caching bugs can silently destroy context windows—test edge cases rigorously
  • Optimizing for cost savings without real-world testing usually costs more in the end

The Bottom Line

Anthropic fessed up, which earns them some street cred. But let's be real—if users hadn't pulled 6,800 session files and compared them, would this have ever gone public? This is a reminder that transparency after exposure isn't the same as proactive honesty. Watch your models like you'd watch your infrastructure—because apparently, nobody else will.