Anthropic is accused of restricting Claude's performance.

The company Anthropic might have intentionally degraded the performance of its models Claude Opus 4.6 and Claude Code? Some developers and AI experts suspect the American company of having capped the performance of its two models. In some cases, these accusations are based on feelings. In others, these allegations are based on studies. On the other hand, Anthropic categorically denies having restricted its two models.

Are Claude Sonnet 4.6 and Claude Opus 4.6 less performant?

On social media, several developers and AI specialists have reported that these two models have become less performant for sustained reasoning, less reliable, and more token-hungry than a few weeks ago. In addition, these two models, widely praised for their performance, would now tend to abandon tasks and hallucinate, while adopting a behavior more focused on the simplest solution.

Through a GitHub discussion, Stella Laurenzo, Principal Director of AMD's AI group, explains that she can no longer simply trust Claude Code for complex engineering tasks. On the other hand, a senior AI executive from a major semiconductor company states that it's not just a feeling and that the regressions are visible in the logs, tool usage patterns, and user corrections.

The performance degradation of Claude Opus 4.6 and Claude Code is therefore not just a matter of feeling, but rather facts, as confirmed by BridgeMind, the company that manages the BridgeBench hallucination test. On April 12, the company recorded that Claude Opus 4.6 dropped from an accuracy of 83.3% (which placed it second in its ranking) to an accuracy of 68.3%, thus relegating it to tenth place in the ranking. For BridgeMind, this is proof that « Claude Opus 4.6 has been braked ». 

In the wake of this, several leading American media outlets took up the story. On social networks, some suspect Anthropic of reducing the performance of its AI to sell them at the same price. Others assume that the company reduced the power of its two models in response to the strong market demand. It is also important to note that other specialists, such as Paul Calcraft, an external researcher in software and AI at X, estimate that these allegations are in some cases exaggerated or based on inadequate evaluation processes.

Anthropic reacts to the accusations

Anthropic refutes the allegations presented above and formally denies any intentional degradation of its models in response to demand pressure. The American company has nonetheless acknowledged making some changes to the usage limits and default reasoning parameters, but without affecting the performance of its models. In response to accusations of throttling, Anthropic has made several changes: introducing adjustable effort parameters, increased transparency in the change log, or displaying implicit reasoning steps to justify the model's responses.

👉🏻 Follow real-time tech news: add 01net to your Google sources, and subscribe to our WhatsApp channel.

Source: VentureBeat