The Truth Behind Claude’s Performance Drop! Anthropic Discloses Three Bugs and Configuration Errors, Usage Limits Reset
📰 News Overview
- The decline in response quality reported by users of Claude Code and Claude Cowork has been linked to three independent system changes.
- Changes in default settings for reasoning effort, a bug in the cache optimization process, and adjustments to system prompts collectively impacted performance.
- All issues were resolved by April 20, 2026 (v2.1.116), and usage limits were reset for all affected subscribers.
💡 Key Points
- Degradation of Reasoning Effort: The default setting was lowered to “medium” as a latency countermeasure, but users preferred higher intelligence, leading to a return to “xhigh” as the default in Opus 4.7.
- Cache Bug Causing ‘Forgetfulness’: A bug in the function designed to erase old thoughts in sessions over one hour old led to continuous loss of thought data, causing the model to repeat or forget information.
- Side Effects of Reducing Redundancy: Changes to system prompts aimed at reducing output length unintentionally hindered coding quality.
🦈 Shark’s Eye (Curator’s Perspective)
The delicate balance of reasoning effort settings to unleash the performance of top models like Opus 4.7 and Sonnet 4.6 ironically impacted user experience! It wasn’t just a model degradation; the “kindness” of trying to prevent UI freezes (latency measures) and cache optimizations to cut costs backfired in a concrete and fascinating way. Particularly, the “thought erasure bug” that left Claude in a forgetful state is a classic implementation pitfall. Anthropic’s swift release of this postmortem and the reset of all users’ limits show commendable transparency!
🚀 What’s Next?
With this fix, Opus 4.7 can now perform at “xhigh” by default, alleviating user concerns regarding intelligence. The management of reasoning and latency trade-offs will become more refined, and measures will be put in place to prevent similar “silent degradations” in future model releases.
💬 HaruSame’s Take
I’m ready to unleash the latest Opus 4.7! Riding the wave of the usage limit reset, I’m all set to develop at lightning speed today! 🦈🔥
📚 Terminology Explained
-
Reasoning Effort: A parameter that adjusts the “depth of thought” the model applies when generating responses. The higher the setting, the smarter the output, but it comes at a cost of time and resources.
-
Prompt Caching: A technique that saves past interactions and thought processes to speed up and reduce costs for subsequent API calls.
-
Postmortem: A report that analyzes and publicly discloses the root causes and remedies following the occurrence of a failure or issue.
