Claude Texts Watermark Possibly Reason for Degraded Quality, Users Report

Anthropic introduced a watermark into all Claude-generated texts starting August 2. Unlike invisible characters or metadata, this is a statistical signal within the words, resilient to copying and editing. The company has not disclosed the mechanics, parameters, or provided a detector. Users have simultaneously noticed that Opus 5, the flagship model, has become harder to work with: answers are too long, convoluted, and hard to parse, with strange vocabulary. The degradation is most pronounced in Claude Code, less so in mobile apps, suggesting a possible link to the watermark that distorts text when there is limited free text (e.g., in code contexts). The author speculates that the watermark forces non-optimal token choices, leading to semantic issues, and notes that Anthropic may have made post-training mistakes. As a result, the author plans to switch to Kimi K3 and Deepseek models, which offer comparable quality at lower cost and are open-source.