Anthropic Claude Performance May Have Been Degraded

Mert Dolay

April 14, 2026

Software developers and artificial intelligence experts have started a major debate on social media in recent weeks; according to allegations, the performance of Anthropic’s flagship models Claude Opus 4.6 and Claude Code has been deliberately downgraded.

Complaints spreading on GitHub, X, and Reddit focus on points such as Claude becoming less capable in complex coding tasks, leaving tasks incomplete, and becoming inefficient in token usage. While some users call this situation “AI shrinkflation,” Anthropic officials strongly deny the allegations that the models were secretly restricted in the name of capacity management.

• Users argue that the logical depth of Claude models has decreased and they make mistakes more frequently.

• Anthropic states that at the root of performance complaints lies not the weakening of the model, but changes in the interface and default settings.

• Researchers state that some viral performance tests contain methodological errors.

• The company confirmed that adjustments were made to session limits to protect user experience during periods of high demand.

Performance Losses Are Defended with Technical Data

A detailed analysis shared on GitHub by a senior executive working at AMD shows that Claude Code has experienced a serious decline since February. The expert, who examined more than 6,000 session files, claims that the model’s logical depth has decreased and its tendency to terminate tasks early has increased.

These technical proofs strengthened the debate by eliminating the feeling of a lack of concrete data beyond personal observations within the community. Technical data turns the disappointment experienced by users into a search for concrete evidence.

Anthropic Explains the Reasons for the Changes

At the center of the allegations lie the statements from the Anthropic side. Company officials emphasize that the “thinking” updates made to Claude Opus 4.6 only aim to improve the user interface and the core capabilities of the model were not touched. It is stated that the changes made to the default “effort level” were implemented to balance token consumption. However, these technical explanations fall short of satisfying users who think the efficiency obtained from the product has decreased.

Lack of transparency leads to trust issues

Silent updates made by the company in its capacity management policies and caching processes arouse suspicion among users that Anthropic is applying secret restrictions. Developers have difficulty understanding whether the performance losses they experience are a software bug or a move aimed at cost-saving.

What do you think, have Claude’s capabilities really been reduced, or do the updates just stem from a misunderstanding? You can share your experiences and views with us in the comments section.