Stella Laurenzo, Director of AI at AMD, recently filed an issue report on GitHub highlighting a significant decline in the performance of Anthropic’s AI coding tool, Claude Code. Laurenzo noted that since February, the model has become noticeably "lazy" and "dim-witted" when handling complex tasks, forcing her engineering team to abandon it.
Laurenzo’s team reached this conclusion after analyzing 6,852 Claude Code sessions, which involved over 230,000 tool calls. The data shows that since the release of Claude Code version 2.1.69 on March 8, there has been a sharp increase in the model failing to think through problems, avoiding tasks, and skipping critical steps. Before the update, such "lazy" behavior was virtually non-existent; now, it occurs an average of 10 times per day.
Lack of Depth in Reasoning Poses Engineering Risks
The report points out that the number of times the model reads files before modifying code has plummeted from an average of 6.6 reads to just 2. Laurenzo believes this is directly linked to the "thinking content redaction" mechanism introduced in the update. By hiding the AI’s thought process by default, the mechanism prevents users from understanding the model’s reasoning logic.
"When the thinking becomes shallow, the model tends to take the path of least resistance: modifying code without reading it, terminating tasks prematurely before they are finished, and opting for the simplest fix rather than the correct one," Laurenzo wrote in the report. She emphasized that all senior engineers on her team have reported similar negative experiences.
To address these issues, Laurenzo has made two specific requests to Anthropic: first, that the platform display the number of "thinking tokens" used per request so users can monitor reasoning depth; and second, that they introduce a "maximum thinking tier" paid subscription for high-level engineering tasks to ensure sufficient computational resources for complex workflows.
Laurenzo’s team has since switched to other AI tools that offer more stable performance. She warned Anthropic that while Claude previously held a lead in reasoning capabilities, the company risks losing its core developer user base if it fails to resolve these performance issues in an increasingly competitive market.
As of press time, Anthropic has not yet responded to a request for comment.
