Anthropic is caught in a growing storm of user complaints claiming its flagship Claude model has become noticeably weaker, less reliable, and more wasteful with computational resources. Over the past several weeks, developers and power users have flooded social media with allegations that Claude Opus 4.6 and Claude Code struggle with sustained reasoning, abandon tasks midway through, and produce more hallucinations than they did just weeks ago . The controversy raises a critical question about how AI companies communicate changes to their models and what happens when inference-time adjustments feel like a silent downgrade to users paying the same price.

What's Actually Changed in Claude's Behavior?

The complaints gained credibility when Stella Laurenzo, Senior Director in AMD's AI group, filed a detailed GitHub analysis on April 2, 2026, backing her claims with hard data . Laurenzo examined 6,852 Claude Code session files, 17,871 thinking blocks, and 234,760 tool calls to document what she saw as a clear regression starting in February. Her analysis showed that Claude's estimated reasoning depth fell sharply while signs of performance degradation rose, including more premature task stopping, more "simplest fix" behavior, and a measurable shift from research-first to edit-first reasoning patterns .

The GitHub post escaped into broader social media discourse when developer Om Patel claimed on X that someone had "actually measured" how much "dumber" Claude had gotten, summarizing the result as a 67% drop . That framing helped popularize the term "AI shrinkflation," the idea that customers are paying the same price for a weaker product. The narrative resonated because it mapped onto what frustrated users reported seeing in practice: more unfinished tasks, more backtracking, more token burn, and a stronger sense that Claude was less willing to reason deeply through complicated coding jobs .

How Did Benchmark Claims Turn Anecdotes Into Proof?

The loudest benchmark-based claim came from BridgeMind, which runs the BridgeBench hallucination benchmark. On April 12, the account posted that Claude Opus 4.6 had fallen from 83.3% accuracy and a number 2 ranking in an earlier result to 68.3% accuracy and number 10 in a new retest, calling that proof that "Claude Opus 4.6 is nerfed" . That post spread widely and became one of the main anchors for the broader public case that Anthropic had degraded the model.

The benchmark claims mattered because they gave the controversy something that looked like hard proof. A developer saying a model "feels worse" is subjective. A screenshot showing a ranking drop from number 2 to number 10, or a dramatic percentage swing in accuracy, gives the appearance of objective evidence, even when the underlying comparison may be more complicated .

However, outside researcher Paul Calcraft challenged the viral comparison on X, arguing that it was misleading because the earlier Opus 4.6 result was based on only six tasks while the later one was based on 30 . On the six tasks the two runs shared in common, Claude's score moved only modestly, from 87.6% previously to 85.4% in the later run, suggesting the dramatic drop was partly an artifact of different test sizes rather than pure model degradation .

What Does Anthropic Say Actually Happened?

Anthropic employees have publicly denied that the company degrades models to manage capacity. Instead, the company has acknowledged real changes to usage limits and reasoning defaults in recent weeks, which has made the broader debate more combustible . Claude Code lead Boris Cherny responded to Laurenzo's GitHub analysis by thanking her for the depth of the analysis but disputing its main conclusion .

Cherny explained that several product changes likely affected what users were seeing. These included:

• UI-only thinking changes: A "redact-thinking-2026-02-12" header that hides thinking from the interface and reduces latency, but does not impact thinking itself, thinking budgets, or how extended reasoning works under the hood .
• Adaptive thinking default: Opus 4.6's move to adaptive thinking by default on February 9, which adjusts reasoning effort based on task complexity .
• Effort level shift: A March 3 shift to medium effort, or effort level 85, as the default for Opus 4.6, which Anthropic viewed as the best balance across intelligence, latency, and cost for most users .

"Users who want more extended reasoning can manually switch effort higher by typing /effort high in Claude Code terminal sessions," stated Boris Cherny, Claude Code lead at Anthropic.

Boris Cherny, Claude Code Lead at Anthropic

The distinction Anthropic is making is technically important but may not satisfy power users who feel the product is delivering worse results. The company is not saying nothing changed. It is saying the biggest recent changes were product and interface choices that affect what users see and how much effort the system expends by default, not a secret downgrade of the underlying model .

Why Does This Matter for Test-Time Compute Strategy?

This controversy highlights a fundamental tension in how AI companies approach inference-time compute, the computational resources spent processing queries after a model is trained. When companies adjust reasoning defaults, effort levels, or thinking visibility, they are making choices about how much compute to allocate per query. These are legitimate product decisions, but they can feel like hidden downgrades to users who expect consistent performance at the same price point .

The Claude situation reveals that transparency about inference-time changes matters as much as the changes themselves. Users can accept trade-offs between speed, cost, and reasoning depth if they understand the trade-off. But when changes feel silent or unexplained, they trigger suspicion and erode trust, even when the underlying model has not degraded .

For enterprises and power users relying on Claude for complex engineering work, extended reasoning is not a luxury but part of what makes the model usable in the first place. When default reasoning effort drops from high to medium, or when thinking is hidden from the interface, the practical experience changes significantly, regardless of whether the model itself has changed .

How to Evaluate Model Performance Changes Yourself

If you rely on AI models for critical work, here are practical steps to assess whether a model has genuinely degraded or whether product changes are affecting your experience:

• Test with consistent parameters: Run the same prompts with identical settings (effort level, thinking budget, context window) across different time periods to isolate whether the model itself has changed or whether product defaults have shifted.
• Document baseline performance: Keep records of model outputs on representative tasks from your workflow, including token usage, reasoning depth, and task completion rates, so you have a historical baseline to compare against.
• Check for product changes: Review release notes and documentation for changes to default settings, reasoning modes, or interface changes that might affect how the model behaves without changing the underlying model weights.
• Evaluate benchmark methodology: When comparing benchmark results over time, verify that the test conditions are identical, including the number of tasks, task difficulty, and evaluation criteria, since different test sizes can produce misleading comparisons.

The Claude controversy underscores a broader lesson for the AI industry: as companies optimize inference-time compute to balance speed, cost, and capability, they need to communicate those trade-offs clearly. Users can adapt to changes they understand. But changes that feel hidden or unexplained, even when technically justified, can spark the kind of trust crisis Anthropic is now navigating .

" }