Users accuse Anthropic of “nerfing” Claude as engineers and executives push back

What's being alleged

It has been reported that a growing number of developers and AI power users are accusing Anthropic of degrading the performance of Claude Opus 4.6 and Claude Code. Complaints have proliferated on GitHub, X and Reddit over the last few weeks — users alleging the models are now worse at sustained reasoning, more likely to stop mid-task, and more prone to hallucinations or contradictions. Some have even coined the term “AI shrinkflation”: same price, less product. Allegedly, a subset of critics think throttling or capacity-driven tuning is to blame.

The concrete complaint driving the firestorm

The story gained traction after a detailed GitHub issue from Stella Laurenzo — who is listed as a Senior Director in AMD’s AI group — analyzed thousands of Claude Code sessions and argued measurable regressions since February: lower estimated reasoning depth, more premature stopping, more reasoning loops and an apparent shift from research-first to edit-first behavior. Those numbers gave the noise a backbone. It has been reported that the post and subsequent screenshots helped the thread go viral, drawing attention from other high-profile users and amplifying frustration into a broader narrative.

Anthropic’s pushback and the murky middle

Anthropic employees have publicly denied that the company intentionally degrades models to manage capacity. At the same time, the company has acknowledged recent changes to usage limits and reasoning defaults — moves that, critics say, could explain altered behavior without being an intentional “nerf.” VentureBeat has reached out to Anthropic for more detail; the company pointed reporters to X posts from Claude Code creator Boris Cherny and team member Thariq Shihipar rather than answering specific questions about inference parameters or benchmark methodology.

Why this matters now

This is about more than model math. It’s about trust. Engineers build pipelines that assume reliable reasoning and tool use; when that expectation slips, workflows break and confidence erodes. So who’s right? Is this a case of users seeing ghosts in the logs, or is the product quietly changing under their feet? Anthropic could calm the waters with transparent benchmarks and clear explanations. Until then, the conversation will stay combustible — and users will keep asking: can we trust Claude today the way we did last month?

Sources: venturebeat.com