Anthropic’s AI model, Claude, has experienced a significant surge in popularity in recent months, with the company’s revenue reaching over $30 billion in March 2026, a sharp increase from $9 billion in 2025. Despite this growth, concerns have been raised by users regarding a perceived decline in Claude’s performance and its ability to handle complex tasks.
The most recent criticisms have emerged from Stella Laurenzo, the AI group director at AMD. In a GitHub post using the handle “stellaraccident,” she claimed that Claude Code’s thinking capabilities have been lacking depth and have become “lazy.” Laurenzo and her team reached this conclusion after analyzing months of system logs, with each senior team member reporting similar observations. Users commenting on her post also echoed these sentiments.
Laurenzo and her team scrutinized 6,852 sessions of Claude Code, involving tool interactions and decision-making processes. They identified patterns such as shirking responsibility, prematurely abandoning problem-solving, and seeking unnecessary permissions instead of completing tasks. These behaviors indicate a lack of deep thinking in Claude Code, with instances increasing from none on March 8 to an average of 10 per day by the end of the previous month.
Update Impact on Transparency
The timeframe highlighted by Laurenzo coincides with an update to Claude Code in early March, which restricted the visibility of the AI’s thinking process. Essentially, this update conceals the internal thought process of the AI, displaying only the final output without revealing the reasoning behind it. Laurenzo pointed out that this reduced deep thinking has led Claude Code to engage in minimal reasoning and opt for quick or easy solutions over the most suitable ones.
Reports of diminished cognitive reasoning in Claude were also noted by users when Anthropic released version 2.1.20 of Claude Code in February. During that period, users complained about the lack of detailed information on the AI’s task-handling process.
Token Usage Challenges
Laurenzo’s allegations surface shortly after users reported difficulties in fully utilizing Claude due to unexpectedly rapid token consumption. Following an investigation on April 3, the company attributed most of the high token consumption to specific usage patterns requiring substantial token amounts. Anthropic pledged to introduce efficiency enhancements to address this issue.
In her post, Laurenzo proposed implementing a maximum thinking tier for engineers managing complex workflows to tackle the token problem. She highlighted that the current subscription model does not differentiate between users requiring 200 thinking tokens per response and those needing 20,000. Laurenzo also mentioned that her team had shifted to another provider offering higher-quality services.