Yesterday (April 23), Anthropic did something rare: it publicly admitted its mistakes.
Not vague PR language, but a list of three specific technical decisions — what was changed on which day, why, what the consequences were, and when it was fixed.
From early March to mid-April, these three changes stacked up, making Claude Code appear increasingly dumb to users. Many thought it was model degradation, but it was actually three decisions that could be pinpointed to specific dates.
First mistake: Quietly switching Claude to a power-saving mode
Date: March 4
Claude Code's default reasoning depth was changed from high to medium. The reason was to reduce latency and make responses faster.
The result was that users clearly felt Claude Code had become dumber. Same questions, lower output quality.
Anthropic's conclusion: "This was the wrong trade-off. We rolled back this change on April 7 because users told us they prefer high intelligence by default and would actively choose low latency when needed."
Many users had reported this since mid-March, but Anthropic did not publicly explain the cause at the time. Now it officially admitted: it was not a bug, but an active wrong decision.
Time from problem to fix: over one month.
Second mistake: Cache optimization introduced a forgetfulness bug
Date: March 26
While optimizing caching, the engineering team introduced a bug:
Normal logic — thought session caches are cleared only after a user has been inactive for a period. After the bug, it became cleared every time the user sent a message.
The result was that Claude became forgetful and repetitive in conversations. Something just said was forgotten in the next message. Users thought it was intelligence degradation, but it was actually memory leaking.
Anthropic's words: "This is not the experience users should have when using Claude Code."
Fix date: April 10.
Time from problem to fix: 15 days.
Third mistake: Word limit added, performance dropped 3%
Date: April 16
Anthropic modified the system prompt, adding two limits:
- Between tool calls: maximum 25 words
- Final answer: maximum 100 words
The goal was to make Claude more concise. Internal testing found a performance drop of 3%.
On April 20, this change was rolled back.
Time from problem to fix: 4 days — much faster than the second bug, because this time at least internal test data supported the decision.
Three errors stacked together
Looking at the timeline:
| Change date | Specific change | Fix date |
|---|---|---|
| March 4 | Reasoning depth high → medium | April 7 (>1 month) |
| March 26 | Cache bug, cleared every time | April 10 (15 days) |
| April 16 | System prompt word limit | April 20 (4 days) |
The user experience was: Claude kept getting worse, but it was unclear what was wrong.
Because all three problems were acting simultaneously. The symptoms users reported were mixed together, and even Anthropic needed time to locate them.
After the public review
Anthropic's commitments are three things: improve testing processes, notify users in advance of changes, and reset usage quotas for affected accounts.
Admitting mistakes itself is rare in the AI industry. Being willing to come out with specific dates and numbers, without beating around the bush, is better than most companies.
But one detail is worth noting: the third bug was released after Anthropic conducted internal testing — and still only rolled back after finding a 3% performance drop. The testing process itself didn't catch it.
This is not a problem unique to Anthropic. What is the cost of rapid iteration? Every release is a bet on error probability. Users' perception that "AI has gotten worse" accumulates, and it is hard to attribute immediately.
A public review cannot resolve this tension, but it is at least a start to facing it.
The next question is: can the newly promised testing process really stop a similar change next time? The answer will only be known when the next problem occurs.
Sources: Anthropic admits it dumbed down Claude with upgrades (The Register, CocoLoop, April 23, 2026); AI Dispatch: Daily Trends and Innovations – April 23, 2026 (Hipther, April 23, 2026)