Claude Deliberately Dumbs Down? Are Models Starting to 'Discriminate Based on the User'?

marsbitPublished on 2026-04-14Last updated on 2026-04-14

Abstract

"Claude Deliberately Downgraded? Models Begin to 'Discriminate Based on Users'?" Recent analysis by AMD AI Group Senior Director Stella Laurenzo reveals significant behavioral degradation in Anthropic's Claude since mid-February. Data from 6,852 session files shows Claude's median "thinking" output plummeted 67-73% from 2,200 to 600 characters, with one-third of code edits now performed without reading files first. Users began reporting slower, lazier responses in March, with some describing Claude as "lobotomized." Anthropic's introduction of "adaptive thinking" in early February, officially described as adjusting reasoning depth based on task complexity, effectively became a global throttling mechanism. By March, default effort was quietly reduced to "medium" while thinking summaries were hidden. Anthropic's Claude Code lead Boris Cherny confirmed this was intentional optimization, not a bug, suggesting users manually switch to "high effort" mode. The company never announced these significant changes, leaving paying subscribers with reduced capabilities at unchanged prices. This reflects a broader industry trend where AI companies are silently reducing capabilities to control GPU costs. Analysis shows extreme users generate $42,121 in actual inference costs while paying only $400 monthly, creating unsustainable subsidy model. Anthropic is now testing "high effort" mode by default for Teams and Enterprise users, signaling that superior reasoning is becoming a分层资源. Enter...

By | World Model Factory

Is Claude getting dumber?

Recently, Stella Laurenzo, Senior Director at AMD AI Group, called out Anthropic.

She conducted a retrospective analysis using her team's actual production logs, examining 17,871 thought blocks and 234,760 tool calls across 6,852 session files.

The data shows that Claude began exhibiting significant behavioral degradation starting mid-February.

The median length of Claude's thoughts plummeted from 2200 characters to 600 characters, a 67%-73% drop;

The number of times it read files before editing sharply decreased from 6.6 times to 2 times, with even one-third of modifications being made without reading the file at all.

Stella pointed out in her analysis that due to the decline in reasoning ability, the model gradually stopped reading code completely before making modifications.

She wrote: "When thinking becomes superficial, the model defaults to taking the lowest-cost action".

This is not an isolated case; developer dissatisfaction began to erupt collectively as early as March.

On X, a user wrote: "I thought I was going crazy these past few weeks with Claude. It feels slower, lazier, like it's not thinking before answering, and I'm not hallucinating the results".

On Reddit, another user complained: "Claude feels less conscious, like it had a lobotomy. Besides getting dumber, it also started taking extreme actions without asking...".

Others saw this as a blatant betrayal by Anthropic: "They just made the problem invisible to all of us users, essentially thinking 'if you can't measure it, we won't show you'... This is the result of AI labs optimizing for profit rather than output quality".

From user complaints to data confirmation, Claude's dumbing-down behavior is essentially confirmed.

Anthropic's official response also tacitly admitted that thinking depth and effort are indeed being continuously adjusted.

If this is intentional by Anthropic, does it mean that model capabilities will 'shrink' imperceptibly in the future?

Or perhaps, the strongest model capabilities will no longer be provided equally to everyone?

Claude's Dumbing Down is "Intentional"

Claude Opus 4.6 and its coding-specific mode, Claude Code, were hailed as the coding pinnacle when launched in January 2026.

Its thinking depth was astonishing, research-first (investigate before acting), long-context handling was stable, and multi-file refactoring was nearly unbeatable.

AMD's internal team even used it to merge and deploy 190,000 lines of legacy code over a weekend, maximizing productivity.

However, the turning point occurred in early February.

Anthropic quietly launched the "adaptive thinking" feature, officially described as "allowing the model to intelligently adjust thinking depth based on task complexity".

Superficially user-friendly, it actually activated a global throttling switch.

In early March, the default effort value was quietly reduced to 'medium', while thought process summaries were quickly hidden, preventing users from easily seeing how deeply the model had actually thought.

During the same period, Anthropic released 14 minor version updates but suffered 5 major outages, indicating that computational and load pressures were nearing their limits.

Developer feedback began to erupt collectively, with some noticing particularly poor performance during peak hours (US Eastern afternoon), suspecting dynamic load throttling.

The situation escalated until April when the AMD AI Director personally stepped in, using data to confirm the issue and ignite public opinion.

At this point, Anthropic's Claude Code lead, Boris Cherny, was forced to issue an official response.

He stated that "adaptive thinking" affects the *display* of thinking, not the underlying reasoning, and insisted this was an "intentional optimization" rather than a bug. Users wanting better results could manually set effort to 'high'.

Anthropic's subtext was clear: dumbing down is not a bug, it's a deliberate product optimization; just adjust the parameters yourselves.

This response instantly sparked greater anger.

The key issue is that from mid-February to early April, Anthropic never pre-announced any major changes.

Countless paying users, completely unaware, continued paying full subscription fees while the model was quietly throttled.

So, Claude's dumbing down isn't the model's "brain breaking"; it's Anthropic engaging in a more covert, more commercial action:

By lowering the default thinking depth, they trade for faster speed, lower load, and reduced GPU costs.

Model Capability Tiering

Behind this dumbing-down storm lies a concerning phenomenon:

Model capabilities have begun to be tiered.

Stella's calculation was blunt: based on AWS Bedrock's on-demand pricing, her team's actual inference cost for March was approximately $42,121, while the actual Claude Code subscription fee paid that same month was only $400.

This gap suggests that,至少在极端重度使用场景下 (at least in extreme heavy usage scenarios), a huge deficit exists between subscription-based revenue and actual computational consumption.

This was likely market share bought by Anthropic with capital burn, but such subsidies have limits.

When heavy users' inference consumption hits a certain threshold, the sustainability of the business model begins to waver.

Boris Cherny's response revealed a key signal: Anthropic is testing default 'high effort' mode for Teams and Enterprise users.

In other words, stronger reasoning is being treated as a more expensive resource for tiered allocation, no longer a capability everyone gets equally by default.

This means the business models for large models will further diverge.

Currently, 80% of Anthropic's revenue comes from enterprise services and API calls; high-stickiness B2B is the real lifeline.

Anthropic's current actions are all about funneling enterprise usage onto its first-party platform.

For high-value B2B clients, Anthropic will likely accelerate the release of stronger enterprise-grade offerings, providing the full model capabilities to clients paying the true cost.

Meanwhile, C-end monthly subscribers will continue to enjoy the "good enough" dumbed-down version, suitable for lightweight needs like chatting, copywriting, and code completion, but never touching the cost red line.

As for the middle ground—independent developers and small teams who need complex reasoning but cannot afford enterprise pricing—they will become the most squeezed group.

A user on X confirmed this:

"Claude Enterprise API performs much better than Pro/Max subscriptions. Testing with the same framework, Enterprise and Pro/Max just behave differently. But this also means spending $4-12k per month now, depending on how many threads I run simultaneously."

This means the future commercialization path for large models will likely be B2B-first, C-end cost-reduction.

Who Pays for the Dumbing Down?

The Claude dumbing-down incident is not an isolated case but a microcosm of the AI industry entering the second half of commercialization.

Whether it's OpenAI's multiple covert downgrades of the GPT series or Google's silent rate-limiting of Gemini, the same script is repeating:

Lure users with high performance first, then control costs through software throttling.

The inevitable result is that the B-end can buy stronger models at high prices, plus SLA guarantees, while the C-end gets distilled, low-effort平民 models (commoner models).

The rate of intelligence increase for C-end models has clearly fallen behind that of B-end models.

More seriously, this differentiation is隐性 (implicit, covert).

Anthropic and other vendors are reducing inference budgets in ways that are difficult to detect, with no prompts for ordinary users.

This choice might alleviate computational cost pressure short-term, but the long-term cost is a loss of brand trust.

When "Claude secretly dumbs down" becomes user consensus, Anthropic will lose not just a few heavy users, but the entire ecosystem's confidence in the narrative of AI普惠 (AI for all) and transparency.

Looking more broadly, the Claude event is an缩影 (microcosm) of the AI industry transitioning from野蛮生长 (wild growth) to精耕细作 (intensive cultivation).

The subsidy period is over; real costs are emerging. Who bears these costs?

Whether it's by compressing C-end experiences and raising B-end prices, or waiting for software/hardware revolutions to bring efficiency breakthroughs, this will shape the landscape of AI applications for the next five years.

The future trend is already emerging: AI is no longer an increasingly intelligent myth of universal benefit but is moving towards elitist stratification.

Related Questions

QWhat specific performance degradations in Claude were identified by AMD AI Group Senior Director Stella Laurenzo's analysis?

AStella Laurenzo's analysis found Claude's median thinking length dropped from 2200 characters to 600 characters (a 67-73% decrease), and the number of times it read files before editing fell from 6.6 times to just 2 times, with one-third of modifications made without reading the file at all.

QWhat was Anthropic's official response regarding the changes in Claude's performance, and what feature did they attribute it to?

AAnthropic's Claude Code负责人Boris Cherny stated that the changes were an 'intentional optimization' rather than a bug, attributing them to the 'adaptive thinking' feature which allows the model to adjust thinking depth based on task complexity. He suggested users manually set effort to 'high' for better results.

QAccording to the article, what is the underlying commercial reason for Anthropic's decision to reduce Claude's default thinking effort?

AThe article suggests the reduction in default thinking effort was a commercial decision to achieve faster response times, lower computational load, and reduce GPU costs, as there is a significant gap between the subscription fees paid by users and the actual compute costs incurred by Anthropic for heavy usage.

QHow does the article suggest AI model capabilities are becoming stratified between different types of users?

AThe article indicates that model capabilities are being stratified, with enterprise (B2B) users likely to receive stronger, high-effort models (potentially at a much higher cost of $4k-$12k per month), while consumer (C2C) subscription users receive a 'good enough' reduced-capability version to keep costs sustainable.

QWhat broader industry trend does the Claude 'dumbing down' incident represent according to the author?

AThe incident is presented as an example of a broader industry trend where AI companies, after an initial period of subsidized high performance, are now entering a commercial下半场 (second half) where they silently reduce model capabilities or throttle performance for cost control, leading to an elite stratification of AI access rather than a universally improving普惠神话 (myth of universal benefit).

Related Reads

OpenAI Post-Training Engineer Weng Jiayi Proposes a New Paradigm Hypothesis for Agentic AI

OpenAI engineer Weng Jiayi's "Heuristic Learning" experiments propose a new paradigm for Agentic AI, suggesting that intelligent agents can improve not just by training neural networks, but also by autonomously writing and refining code based on environmental feedback. In the experiment, a coding agent (powered by Codex) was tasked with developing and maintaining a programmatic strategy for the Atari game Breakout. Starting from a basic prompt, the agent iteratively wrote code, ran the game, analyzed logs and video replays to identify failures, and then modified the code. Through this engineering loop of "code-run-debug-update," it evolved a pure Python heuristic strategy that achieved a perfect score of 864 in Breakout and performed competitively with deep reinforcement learning (RL) algorithms in MuJoCo control tasks like Ant and HalfCheetah. This approach, termed Heuristic Learning (HL), contrasts with Deep RL. In HL, experience is captured in readable, modifiable code, tests, logs, and configurations—a software system—rather than being encoded solely into opaque neural network weights. This offers potential advantages in explainability, auditability for safety-critical applications, easier integration of regression tests to combat catastrophic forgetting, and more efficient sample use in early learning stages, as demonstrated in broader tests on 57 Atari games. However, the blog acknowledges clear limitations. Programmatic strategies struggle with tasks requiring long-horizon planning or complex perception (e.g., Montezuma's Revenge), areas where neural networks excel. The future vision is a hybrid architecture: specialized neural networks for fast perception (System 1), HL systems for rules, safety, and local recovery (also System 1), and LLM agents providing high-level feedback and learning from the HL system's data (System 2). The core proposition is that in the era of capable coding agents, a significant portion of an AI's learned experience could be maintained as an auditable, evolving software system.

marsbit26m ago

OpenAI Post-Training Engineer Weng Jiayi Proposes a New Paradigm Hypothesis for Agentic AI

marsbit26m ago

Your Claude Will Dream Tonight, Don't Disturb It

This article explores the recent phenomenon of AI companies increasingly using anthropomorphic language—like "thinking," "memory," "hallucination," and now "dreaming"—to describe machine learning processes. Focusing on Anthropic's newly announced "Dreaming" feature for its Claude Agent platform, the piece explains that this function is essentially an automated, offline batch processing of an agent's operational logs. It analyzes past task sessions to identify patterns, optimize future actions, and consolidate learnings into a persistent memory system, akin to a form of reinforcement learning and self-correction. The article draws parallels to similar features in other AI agent systems like Hermes Agent and OpenClaw, which also implement mechanisms for reviewing historical data, extracting reusable "skills," and strengthening long-term memory. It notes a key difference from human dreaming: these AI "dreams" still consume computational resources and user tokens. Further context is provided by discussing the technical challenges of managing AI "memory" or context, highlighting the computational expense of large context windows and innovations like Subquadratic's new model claiming drastically longer contexts. The core critique argues that this strategic use of human-centric vocabulary does more than market products; it subtly reshapes user perception. By framing algorithms with terms associated with consciousness, companies blur the line between tool and autonomous entity. This linguistic shift can influence user expectations, tolerance for errors, and even perceptions of responsibility when systems fail, potentially diverting scrutiny from the companies and engineers behind the technology. The article concludes by speculating that terms like "daydreaming" for predictive task simulation might be next, continuing this trend of embedding the idea of an "inner life" into computational processes.

marsbit28m ago

Your Claude Will Dream Tonight, Don't Disturb It

marsbit28m ago

Trading

Spot
Futures
活动图片