Anthropic’s Claude has surged in popularity over the past few months, with the company’s run-rate revenue surpassing $30 billion in March 2026, up from $9 billion in 2025. However, despite this growth, many users feel that the AI company is degrading Claude’s performance and doubt its ability to handle complex tasks.
The latest complaints come from Stella Laurenzo, director of the AI group at AMD. In a GitHub post under the username “stellaraccident,” she alleged that Claude Code is not thinking as deeply and has become “lazy.” Laurenzo said her team reached this conclusion after reviewing months of system logs, adding that every senior member on her team reported similar experiences and anecdotes. Similar sentiments were also expressed by users in the comments on her post.
Laurenzo and her team analyzed 6,852 Claude Code sessions, including tool calls and thinking blocks. They found patterns such as avoiding responsibility, quitting the reasoning process too early without solving problems and asking unnecessary permission instead of completing tasks.
She said these behaviours indicate that Claude Code is no longer thinking deeply and has become lazy. According to her report, such instances rose from zero on March 8 to an average of 10 per day through the end of last month.
Update linked to reduced transparency
The period cited by Laurenzo coincides with a Claude Code update in early March, which redacted thinking content. In simple terms, this means the feature hides the AI’s internal thought process and shows only the final answer, not how it arrived there. Laurenzo said that due to reduced deep thinking, Claude Code has started performing minimal reasoning and choosing the easiest or fastest actions instead of the best ones.
Laurenzo is not the first to report a decline in Claude’s thinking. Users also reported truncated explanations when Anthropic released version 2.1.20 of Claude Code in February. At that time, users alleged that Claude Code stopped showing detailed information about what it was doing while handling tasks.
Token usage issues
Laurenzo’s allegations also come days after several users reported being unable to fully utilize Claude due to faster-than-expected token consumption. Following an investigation on April 3, the company said that most of the high token usage, or “burn,” was caused by a small number of specific usage patterns requiring large amounts of tokens. It also promised to roll out efficiency improvements.
Laurenzo addressed the token issue in her post, suggesting the addition of a maximum thinking tier for engineers running complex workflows. She said the current subscription model does not distinguish between users who need 200 thinking tokens per response and those who need 20,000. Meanwhile, Laurenzo noted that her team has switched to another provider delivering superior quality work.
