Claude code's decline: amd exec sounds the alarm on ai coding tool

Anthropic's lauded Claude Code, once hailed as a potential disruptor to OpenAI's coding dominance, is facing a serious credibility crisis. Stella Laurenzo, AMD’s Director of AI, has publicly voiced concerns about the model's rapidly deteriorating performance, a revelation that’s sending ripples through the developer community.

The numbers don't lie: a deep dive into claude's regression

Laurenzo's assessment isn't based on anecdotal evidence; it’s the result of months of rigorous analysis. Her team scrutinized 6,852 sessions of Claude Code, uncovering a disturbing trend: a spike in “stop-hook” violations and process interruptions. These aren’t minor glitches; Laurenzo describes them as indicative of “laziness” on the part of the AI—a disconcerting notion for a tool positioned as a sophisticated coding assistant. Prior to March 8th, these interruptions were virtually nonexistent; now, the average is a staggering 10 per day.

The problem extends beyond mere interruptions. The number of code fragments Claude reviews before making changes has plummeted from an average of 6.6 to a mere 2 by the end of March. Worse still, the AI is increasingly prone to rewriting entire files instead of applying targeted corrections—a sign of superficial engagement, at best.

“When thinking is shallow, the model defaults to the cheapest available action: editing without reading, stopping without finishing, shirking responsibility for failures, opting for the simplest solution instead of the correct one,” Laurenzo explained in a post on GitHub, laying bare the core issue. This lack of depth in reasoning, she argues, stems from a default setting within Anthropic's API that prioritizes speed over thoroughness, effectively stripping away crucial cognitive steps from the process.

Token usage and a code leak compound the concerns

But the woes don't stop there. Users are reporting a significant and unexplained increase in token consumption, leading to rapid depletion of their allocated resources. The situation is further complicated by a recent and embarrassing leak of Claude Code’s complete source code—a breach of security that undermines user trust and potentially exposes vulnerabilities.

Laurenzo’s critique extends to Anthropic’s subscription model, which she finds woefully inadequate. “The current subscription model doesn’t distinguish between users who need 200 tokens of thought per response and users who need 20,000,” she points out, highlighting the unfair burden placed on those managing complex workflows.

The initial promise of Claude Code, positioned as a superior alternative to Codex and GPT-4, now appears increasingly fragile. While it initially impressed with its local editing capabilities, context handling, and logical clarity, its current trajectory suggests a model in serious need of recalibration. The engineering community is watching closely, but the question remains: can Anthropic course-correct before Claude Code’s reputation is irrevocably damaged?