Claude performance degradation has become one of the most talked about topics in AI circles over the past few weeks, and the frustration is hard to ignore. I actually have seen some responses in my Claude which made me raise an eyebrow. It was a little poor in quality compared to how it was days ago. Turns out, I wasnt imagining things.
A wave of Claude power users is accusing Anthropic of quietly making Claude Opus 4.6 and Claude Code worse. The complaints are consistent: more abandoned tasks, higher token burn, shallower reasoning, and more hallucinations than people remember from a few weeks ago. The controversy picked up serious traction after an AMD AI senior director posted a detailed GitHub analysis of thousands of Claude Code sessions, arguing the tool has regressed for complex engineering work.
What Anthropic Actually Said
Anthropic’s response is essentially: yes, things changed, but not the core model in the way people are alleging. Company staff and Claude Code lead Boris Cherny pointed to recent updates being mostly product defaults and UI behavior changes. Examples include adaptive thinking and medium effort becoming the new default, plus usage limit adjustments to manage demand spikes.
Critics are not buying it. Whatever the root cause, the lived experience feels worse. “AI shrinkflation” has become the popular framing, and honestly it is understandable. You are paying the same price and getting output that feels less reliable than before.
Why This Matters Beyond the Drama
Developer and power user trust is fragile. Once advanced users start assuming model quality is variable or capacity-throttled, they will build their workflows, budgets, and vendor decisions around that uncertainty. That is a slow bleed for Anthropic even if the underlying model has not changed at all.
It also highlights a genuine problem with how AI products are built right now. Separating “the model got worse” from “the defaults changed” is genuinely difficult when the product surface keeps moving. Effort levels, reasoning visibility, rate limits, UI behavior. Any of these can shift your results without touching a single weight in the model.
What You Should Do If You Rely on Claude
If Claude is part of your actual workflow, here are three practical steps worth taking now.
First, lock down your settings. Do not rely on platform defaults. Set your effort level explicitly, keep your system prompts consistent, and document what your baseline looks like.
Second, build your own simple regression tests. You do not need anything fancy. A handful of prompts you run regularly and compare against previous outputs will tell you faster than any forum post whether something has actually changed for your use case.
Third, diversify where it makes sense. Claude is still excellent. But if a single model is a single point of failure in something important, that is worth addressing regardless of this particular situation.
The Bigger Picture
This story is a good reminder that AI tools are not static products. They are living systems with changing defaults, capacity constraints, and ongoing tuning. That is not inherently bad, but it does mean your experience today is not guaranteed to be your experience next month.
Anthropic will likely need to be more transparent about what changes when and why, especially for users paying for high-tier access. The alternative is a community that assumes the worst every time something feels off, and that is not good for anyone.
For now, treat your AI setup the way you would treat any critical tool. Verify it. Test it. Do not assume consistency just because the interface looks the same. This sure is going to be an interesting development.