AMD Critiques Anthropic's Claude Code: Reliability in Question

AMD Critiques Claude Code from Anthropic Severely

Since February, Claude Code, Anthropic's flagship coding tool, has come under fire. The company AMD, through the voice of Stella Laurenzo, head of the AI group, expressed its dissatisfaction via a ticket published on GitHub. This ticket summarizes an in-depth analysis conducted by her team, which examined 6,852 sessions of Claude Code, totaling 234,760 tool calls and 17,871 reasoning blocks. The conclusion of this study is unequivocal: Claude Code can no longer be considered reliable for complex engineering tasks.

A Controversial Update

AMD's analysis points to a specific update of Claude Code. Since the deployment of version 2.1.69 in early March, Anthropic has enabled by default a feature called "thinking redaction." This mechanism has the effect of masking the content of internal reasoning in the API responses. Although reflection tokens continue to be generated, users can no longer see the AI's thought process before it acts.

According to data collected by AMD, this modification has led to a decrease in reasoning depth. As a result, Claude Code tends to take shortcuts, such as modifying code without reviewing it, stopping before completing a task, or evading errors instead of correcting them. This observation has been confirmed by every senior engineer on the AMD team, and the GitHub ticket has sparked numerous similar testimonials, both on the platform and on Reddit.

AMD's Requests to Anthropic

In light of this situation, Stella Laurenzo made several requests to Anthropic. First, she demands complete transparency regarding any potential capping of reasoning tokens. Next, she wants the number of reflection tokens used per request to be disclosed, so users can monitor the depth of processing. Finally, she proposes the creation of a subscription tier that includes longer reasoning for engineering teams. Currently, the model does not differentiate between a user consuming 200 reflection tokens per response and another who needs 20,000.

A Significant Defection

In response to these issues, AMD's AI team has decided to turn to a competitor, whose name remains confidential due to a non-disclosure agreement. In an industry where even Microsoft is testing Claude Code internally, losing a client of AMD's stature sends a strong signal. Stella Laurenzo warns that Anthropic is no longer alone at the top.

A Difficult Context for Anthropic

This criticism from AMD comes at a time already tense for Anthropic. On March 31, a massive leak of Claude Code's source code occurred, exposing 512,000 lines of TypeScript via a source map file mistakenly published on the npm registry. In attempting to clean up this leak, Anthropic deleted 8,100 repositories on GitHub, including legitimate forks, causing an uproar among developers.

In February, the company had already cut access to third-party applications using Pro and Max subscriptions, alienating part of its developer community. This series of problems arises as Claude Code's user base experiences explosive growth. After discussions with the Pentagon regarding the military use of Claude, daily active users surged from 4 million in January to 11.3 million in early March, with one million sign-ups per day for an entire week. Web traffic skyrocketed by 297% in a year, making Claude the most downloaded application in the United States and 15 other countries, including France.

The Consequences of Growing Popularity

This popularity comes at a cost. Outages have multiplied, with major incidents on March 25, 26, and 27, followed by March 31, April 1, 3, and 6. On April 6, over 8,000 users reported issues on Downdetector. The uptime of claude.ai in March fell to 98.21%, well below the standards expected for a professional service.

The Pro subscriptions ($20 per month) and Max subscriptions ($100 to $200) are massively subsidized compared to the actual cost of tokens consumed via the API. An intensive developer on Claude Code can burn through the equivalent of several hundred dollars in API tokens in a single day while only paying a fixed monthly fee. This business model works as long as the user base remains manageable. But when it triples in two months, the economic equation becomes untenable.

AMD's ticket raises a crucial question without explicitly stating it. Is Anthropic throttling reasoning to absorb the growing load? Or is the degradation of Claude Code a collateral damage of rushed updates? With subscriptions too cheap and an API too expensive for independents, the middle ground remains elusive, and it is the quality of service that suffers.