Anthropic Confirms Claude’s Capability Reduced by Recent ‘Upgrades’

Anthropic Confirms Claude’s Capability Reduced by Recent ‘Upgrades’

An investigation by Anthropic has confirmed recent degradation in Claude’s performance, affecting user experience across several features. Over the past month, customers reported a decline in the quality of responses, particularly with Claude Code, the Claude Agent SDK, and Claude Cowork. These changes coincided with three significant adjustments made in March and April, although the Claude API remained unaffected.

Recent Changes Impacting Claude’s Performance

  • Adjustment of Reasoning Effort: On March 4, Anthropic altered the default reasoning effort for Claude Code from high to medium. This change aimed to reduce latency but inadvertently lowered the model’s reasoning capability. After consistent user feedback, the default was reverted on April 7.
  • Cache Optimization Bug: On March 26, a bug emerged following a cache optimization effort. This bug led to the clearing of all cached session data with every prompt response. As a result, users faced issues with Claude becoming forgetful and repetitive. This problem was addressed on April 10.
  • System Prompt Revision: A system prompt adjustment on April 16 aimed to reduce verbosity. However, subsequent testing indicated a 3% drop in performance for several models. This change was reverted on April 20.

Future Improvements and Commitments

In response to these issues, Anthropic has committed to conducting more rigorous internal testing for upcoming releases of Claude Code. The company also plans enhancements to its Code Review tool and will improve the assessment of system prompt modifications.

To facilitate better communication, Anthropic launched a new @ClaudeDevs account on X. This initiative aims to provide users with deeper insights into product decisions and the rationale behind them. Recently, the head of growth, Amol Avasare, emphasized the importance of direct communication with users to address concerns more effectively.

Responding to user dissatisfaction, Anthropic has reset account usage levels to ensure a more favorable experience. The company acknowledged that the recent performance was below expectations for Claude Code.

Next