Anthropic has officially acknowledged what many users suspected: Claude was inadvertently “dumbed down” in recent weeks due to multiple technical issues.
**The Three Bugs**
In a postmortem published on their blog, Anthropic revealed three interconnected issues:
1. **Reasoning Level Reduced**: The reasoning level was quietly changed from “high” to “medium,” affecting the depth of analysis Claude could provide.
2. **Cache Bug**: A bug in the caching system was clearing conversation history mid-session, causing Claude to lose track of ongoing discussions.
3. **25-Word Limit**: A prompt limit was affecting output quality, causing responses to be truncated or simplified inappropriately.
**Impact on Users**
The issues combined to create a noticeably degraded experience. BridgeMind’s BridgeBench showed Opus 4.6 accuracy dropping from 83.3% to 68.3%, with the model falling from second place to tenth in rankings.
Users reported experiencing:
– Shorter, less detailed responses
– Repeated questions about information already provided
– Difficulty with complex, multi-step tasks
– General decline in reasoning quality
**The Timing Question**
The postmortem was published shortly after OpenAI released GPT-5.5, leading some observers to question whether the timing was coincidental. Anthropic CEO Dario Amodei has not commented publicly on the timing.
**What’s Been Fixed**
Anthropic states that all three issues have been resolved:
– Reasoning level has been restored to “high”
– Cache bug has been patched
– Word limit has been removed
However, the company notes that “prompts written for earlier models may sometimes produce unexpected results” with the restored capabilities—meaning users may need to adjust their prompting style.
**Lessons Learned**
The incident highlights the complexity of maintaining large language models at scale. Even established companies like Anthropic can experience quality regressions that significantly impact user experience.
For users, the episode serves as a reminder: AI capabilities aren’t static, and occasional regressions are part of working with rapidly evolving technology.
—
*Did you notice changes in Claude’s performance recently? Share your experience.*
Leave a Reply