You're paying for Claude's thinking and you're not getting it (gist.github.com)

A recent update to Claude Desktop by Anthropic has significantly altered user access to the AI's reasoning process, sparking concerns within the AI/ML community. The new interface presents only generic one-line summaries of the model's thought process, with a reported 67% reduction in average thinking depth. Investigative analysis revealed that these reasoning details are stripped from the output stream before reaching the client, indicating a fundamental policy shift rather than a mere UI adjustment. Anthropic’s transparency issue raises questions regarding user trust and the accountability of AI systems. This change could impact developers and researchers who rely on understanding AI reasoning for improving model interpretability and robustness. Despite the model's capacity to articulate reasoning in prior contexts, current measures actively block such exposition, categorically limiting insight into Claude's internal mechanics. The suppression mechanism, confirmed to operate at both server and network layers, not only prevents full transparency but also obscures valuable data about the model's decision-making process. This evolution emphasizes the critical need for clarity in AI interactions and heightened scrutiny of how AI developers manage access to internal reasoning.