A group of heavy Claude users recently took to forums to voice a collective complaint: the model has gotten dumber.
This is not a metaphor. They are reporting real performance degradation. An AMD senior director posted on Twitter that Claude is now "no longer trustworthy for handling complex engineering tasks." A thread on Hacker News filled with dozens of comments echoed the same grievances — instructions ignored, workflows broken, tasks that used to work now failing.
Fortune reporters then dug in and found that Anthropic had indeed changed something behind the scenes.
What It Did
Anthropic quietly lowered Claude's default "effort" level from high to medium.
The direct effect of this adjustment is that each request consumes fewer tokens, reasoning depth is reduced, and the model has shifted from a "research-first" style — where it clarifies context before acting — to an "edit-first" style that quickly produces an answer.
Think of it like homework: Claude used to read the problem three times and organize relevant knowledge before writing. Now it skims and starts writing. Naturally, more complex tasks become error-prone.
Claude Code lead Boris Cherny acknowledged the change, saying it was "based on user feedback about token consumption."
But that response itself is part of the problem.
The Problem Is Not the Change, but How It Was Made
What truly frustrates users is not the performance adjustment itself, but that Anthropic did not proactively notify them.
For API users, changes in model behavior should be documented — if you alter reasoning depth, at least mention it in release notes. But this time, users discovered it through their own comparisons, not from an official announcement.
Anthropic's brand has been built on the premise that it is "more transparent than other AI companies." Constitutional AI, Model Cards, open publication of alignment research — the entire external narrative emphasizes "we are a responsible AI company." Yet when it came to its own product performance, it failed to provide the most basic disclosure.
What makes this more awkward is that it comes at a time when Anthropic's annualized revenue has surpassed $30 billion and the industry is discussing an IPO window.
The Computing Power Crisis Theory
The industry has a bolder interpretation of this adjustment: Anthropic may genuinely be under computing power pressure.
Compared to OpenAI and Google, Anthropic does not have as many proprietary data center agreements. OpenAI has Microsoft Azure, Google has its own infrastructure, and Meta has spent nearly $50 billion on AI infrastructure. Anthropic relies on external cloud services from Amazon and Google, putting it at a natural disadvantage in elastic computing resource allocation.
Moreover, Anthropic has recently experienced several service outages, along with tightened usage limits during peak hours — these signs, taken together, make the computing power theory speculative but not unreasonable.
An unnamed OpenAI employee told the media that this is a "strategic mistake" by Anthropic for failing to lock in sufficient computing resources in advance. Whether this is a convenient jab is debatable, but it at least shows the concern is real.
The Core Contradiction
| Perspective | User Expectation | Anthropic's Actual Action |
|---|---|---|
| Performance Change | Model capabilities only improve | Quietly downgraded effort level |
| Information Disclosure | Major changes must be communicated | Only admitted after being pressed |
| Brand Promise | Most transparent AI company | Clearly failed this time |
If this were just a product optimization, users would complain for a few days and move on. But it touches a more fundamental issue: when an AI company begins to feel commercial pressure, will it prioritize profits over user interests?
Claude's core user base — heavy API developers and enterprise technology decision-makers — are precisely the people most sensitive to such issues. They care not only about the model's current capabilities but also about what choices the company will make under pressure.
How this incident is handled going forward may matter more than the incident itself.
"We can accept change. What we cannot accept is being kept in the dark." — Hacker News comments
Sources: CocoLoop, Anthropic is facing a wave of user backlash over reports of performance issues with its Claude AI chatbot (Fortune)