Anthropic Faces User Backlash Over Reported Performance Degradation in Claude AI Models

Anthropic faces user backlash over degraded performance in its Claude AI models. Users, including an AMD director, report more errors and failures on complex tasks, labeling it "AI shrinkflation." Anthropic denies "nerfing" its models, attributing changes to new default "effort" settings designed to save tokens. The controversy is fueled by speculation about a compute capacity crunch at the rapidly growing company, which has also experienced recent outages.

Claude performance degradation complaints

• ▪User complaints about Claude's performance have spread on platforms like GitHub, X, and Reddit.
• ▪Some users have termed the perceived issue "AI shrinkflation," suggesting customers are paying the same price for a weaker product.
• ▪Laurenzo's analysis of over 6,800 session files found that Claude shifted from a "research-first" approach to a faster but more error-prone "edit-first" style.
• ▪A widely shared GitHub analysis by Stella Laurenzo, a senior director of AI at AMD, concluded that Claude Code had become "unusable for complex engineering tasks."
• ▪Developers and users report Anthropic's Claude AI models are failing to follow instructions, making more mistakes, and struggling with complex workflows.

Reasoning defaults product changes

• ▪A GitHub issue alleged Claude Code's prompt-cache time-to-live (TTL) was changed from one hour to five minutes in early March 2026, increasing token usage for some users.
• ▪While Claude Code users can manually adjust effort levels, users of the Pro versions of Cowork or the desktop Claude cannot change the default setting.
• ▪In early February 2026, Claude Opus 4.6 introduced "adaptive thinking," allowing the model to decide how much reasoning to apply to a task.
• ▪Anthropic changed the default "effort" level for its Claude model to "medium" to economize on the number of tokens processed per request.

Benchmark methodology disputes

• ▪Critics argued the BridgeBench comparison was misleading, as the first result was based on six tasks while the second was based on 30.
• ▪A viral post by BridgeMind claimed Claude Opus 4.6's accuracy on its BridgeBench benchmark dropped from 83.3% (rank #2) to 68.3% (rank #10).
• ▪Another viral post on X by developer Om Patel on April 7, 2026, claimed a measured 67% drop in Claude's capability.
• ▪On the six tasks common to both BridgeBench runs, Claude's score changed only slightly, from 87.6% to 85.4% accuracy.

Session capacity limits

• ▪On April 15, 2026, Anthropic's Claude services experienced a widespread outage that lasted several hours.
• ▪A leaked memo from competitor OpenAI claimed Anthropic made a "strategic misstep" by not securing enough compute capacity.
• ▪On March 26, 2026, Anthropic adjusted 5-hour session limits for some subscribers during peak hours to manage growing demand.
• ▪Anthropic has trained a new, more capable model called Mythos, which is larger and more expensive to run, but has not been released to the public.

Anthropic public denials

• ▪An Anthropic spokesperson publicly stated that the company does not degrade its models to manage demand.
• ▪Boris Cherny, head of Claude Code, called accusations that Anthropic secretly "nerfed" its model "false."
• ▪Cherny stated the default "effort" change was made in response to user feedback about high token consumption and was noted in the changelog.