Claude翻车：Opus 4.1白天退化，Anthropic承认并回滚更新

Core Viewpoint - Claude Opus 4.1 has experienced performance degradation, particularly during daytime usage, leading to user complaints about its responsiveness and accuracy [1][2][14]. Performance Issues - Users reported that Claude Opus 4.1's reasoning performance significantly declined between 10 AM and 11 AM, with many errors occurring during document processing tasks [2][3]. - The performance degradation appears to resolve during nighttime, indicating a potential issue with the model's operational parameters during peak usage hours [3]. Technical Aspects - Speculation suggests that the performance issues may be linked to the use of 1.58-bit quantization, which significantly impacts model precision [4][5]. - This quantization method reduces model parameters to only three values {-1, 0, 1}, which can lead to a loss of critical information and affect the model's ability to handle complex tasks effectively [5][7]. - The limitations of 1.58-bit quantization may also compromise the model's stability, particularly in applications requiring precise data handling, such as medical image analysis and financial risk prediction [8]. User Experience and Limitations - Users have reported reaching usage limits within two hours, with inconsistent responses from customer service regarding whether the limits are based on usage time or volume [9][10]. - There are also reports of the model exposing API keys, raising concerns about security and reliability [12]. Company Response - Anthropic acknowledged the issues with Claude Opus 4.1, attributing the performance decline to problems in the reasoning stack that were intended to enhance model efficiency [14]. - The company has rolled back to a previous version, Claude Opus 4.0, which also faced similar issues, demonstrating a proactive approach to addressing user concerns [14][17].