Anthropic's Claude Models Face Widespread Performance Regression, Sparking Developer Outcry

Widespread reports indicate a significant performance regression across Anthropic’s Claude models, particularly Opus 4.7 and Sonnet 4.6, with developers describing them as ‘dumber and lazier’ since recent updates. A quantitative analysis by AMD’s AI director highlights a direct correlation between ‘thinking content redaction’ and a measured quality decline in complex engineering workflows, noting a 73% drop in thinking depth and significant increases in user frustration and reasoning loops. This degradation is further substantiated by Margin Labs, which reported a consistent dip in SWE bench performance from 57% to 55% since March, and BridgeMind’s hallucination benchmark showing Opus 4.6 regressing from 87.6% to 73.3% from launch to April 12th, validating anecdotal user experiences of task refusals, ‘dumber solutions,’ and ‘getting lost.’

Technical scrutiny reveals multi-layered contributing factors, including critical inefficiencies within the Claude Code harness, which benchmarks show performing 15% worse than alternative interfaces like Cursor and scoring merely 58% on Terminal Bench. API-level changes are also implicated, such as Opus 4.7’s updated tokenizer increasing token counts by up to 1.47x for the same content, potentially inducing context bloat. Furthermore, Anthropic’s September post-mortem acknowledged that the 1 million token context window version of the model is ‘dumber,’ yet this version is now the default for Claude Code users, requiring manual opt-out. These issues collectively contribute to a drastic increase in resource consumption, with 80 times more API requests and 64 times more output tokens yielding demonstrably worse results for the same human effort, a stark contrast to the consistent stability observed in OpenAI’s models.