There was a moment when you typed a prompt into Claude and felt like... it was answering like an intern on their first day.
You were debugging when you noticed Claude suggesting strange code. You kept fixing it, but it wouldn't work. Or worse, it kept suggesting the exact same solution over and over, as if it didn't know you just rejected it.
The first time, you blamed your prompt. The second time, you tried rephrasing. The third time, you started wondering... is it me, or is it the AI?
Anthropic just released a post-mortem that answers exactly that.
From Reddit Complaints to Official Admission: When Customer Voices Are Loud Enough for Anthropic to Listen
Before Anthropic's official post-mortem, the first signs came from everyday users on Reddit and X complaining that Claude had changed. It wasn't just a feeling—someone even went as far as pulling over 6,800 session files to analyze, finding clear evidence of a quality drop. This is a great reminder that your customers will usually feel the problem long before your monitoring systems detect it. How about your business? Do you have channels for your customers to tell you this?
Mistake 1: Trading "Depth" for Less Screen Lag—Without Telling Users
Anthropic tried to solve screen latency by quietly reducing the Reasoning Effort from High to Medium without any announcement. As a result, the AI did respond faster, but it chose "the easiest way out" instead of "the most accurate solution." Imagine if your clients were using Claude to analyze critical data during that time—the answers they received might not have been the best ones, and they wouldn't have even realized it.
Mistake 2: The Bug That Made Claude "Forget" Everything Mid-Conversation
There was a minor bug in the caching system intended to run only once when a session expired, but instead, it ran every single time. The result? Claude forgot the entire context of the conversation every time a new turn started. If you felt like you had to explain yourself repeatedly in the same chat, or if it kept looping back to the same answers as if it didn't hear you, that was this bug at work.
Mistake 3: Forcing Shorter Answers Led to Missing Information
Anthropic's team added a system prompt to reduce verbosity and save tokens. The side effect was that Claude opted to cut out necessary reasoning steps and crucial details. Output quality, especially for coding, dropped measurably. This serves as a strong reminder: optimizing to save costs without testing the impact on real users usually ends up costing more in the long run.
The Good News: Everything Is Fixed, and Anthropic Refunded Usage Limits
As of April 20, 2026, all 3 issues were completely fixed in patch v2.1.116. The Reasoning Effort is back to High, with Opus 4.7 specifically set to 'xhigh' as compensation. The caching bug was squashed, and the limiting prompt was removed. Furthermore, Anthropic refunded Usage Limits to all subscribers starting April 23 as a tangible apology. What's more interesting than the bug fixes is that they chose to come out and tell the complete truth—something not every company is brave enough to do.
Conclusion
These 3 problems didn't happen because the AI devolved, but because of decisions made by humans trying to improve it and slightly missing the mark. The real-world lesson for every business here is that a seemingly minor system change can unintentionally have a massive impact on the customer experience. Most importantly, transparency when you mess up is what truly buys long-term trust.
Originally published at ireadcustomer.com.
Top comments (0)