For several weeks, the AI developer community has been sounding the alarm over a perceived decline in the intelligence of Anthropic’s flagship models. Users across platforms like GitHub, X, and Reddit reported a phenomenon they dubbed “AI shrinkflation” —a trend where Claude appeared less capable of complex reasoning, more prone to errors, and increasingly inefficient with token usage.
While Anthropic initially denied claims that they were intentionally “nerfing” the model to manage server demand, the company has now released a technical post-mortem. The investigation reveals that while the core AI models remained unchanged, three specific adjustments to the “harness”—the software layers surrounding the model—unintentionally crippled its performance.
The Evidence of Decline
The backlash was not merely anecdotal; it was backed by significant technical data. In early April 2026, the narrative of a “dumbing down” of Claude gained mainstream momentum through several key findings:
- Large-scale Audits: Stella Laurenzo, Senior Director at AMD’s AI group, conducted an audit of over 6,800 Claude Code sessions and 234,000 tool calls. Her data suggested a sharp decline in reasoning depth, noting that the model often fell into repetitive loops or opted for the easiest possible fix rather than the most accurate one.
- Benchmark Drops: Third-party testing by BridgeMind showed a significant hit to Claude Opus 4.6, with accuracy scores reportedly falling from 83.3% to 68.3%, causing its industry ranking to drop from second to tenth place.
- Resource Inefficiency: Users reported that usage limits were being exhausted much faster than usual, fueling suspicions that the model was becoming “wordier” or less efficient in how it processed information.
Why It Happened: Three Technical Culprits
Anthropic clarified that the “brain” of the AI (the model weights) had not changed. Instead, the issues stemmed from changes made to the user interface and the instructions guiding the model’s behavior:
1. Reduced Reasoning Effort
To solve UI latency issues—where the interface appeared “frozen” while the model “thought”—Anthropic lowered the default reasoning effort from high to medium for Claude Code. While this made the interface feel faster, it stripped the model of the computational depth required for complex engineering tasks.
2. The Caching Logic Bug
A March 26 update intended to optimize memory by pruning old “thinking” data from idle sessions contained a critical error. Instead of clearing old data once after an hour of inactivity, the bug cleared the model’s “short-term memory” during every subsequent interaction. This caused Claude to become forgetful and repetitive.
3. Verbosity Constraints
In an attempt to make responses more concise, Anthropic introduced new instructions to limit the length of text between tool calls and final responses. This “brevity” mandate backfired, with evaluations showing a 3% drop in coding quality as the model struggled to express complex logic within strict word counts.
Restoring Trust and Future Safeguards
The impact of these errors was felt across the Claude Code CLI, the Claude Agent SDK, and Claude Cowork, though the core Claude API remained unaffected. To rectify the situation and prevent a recurrence, Anthropic is implementing a series of structural changes:
- Expanded “Dogfooding”: More internal staff will be required to use the exact same public builds as customers to catch regressions before they reach the market.
- Rigorous Testing: The company is deploying enhanced evaluation suites to test how every minor change to a “system prompt” affects the model’s overall intelligence.
- Subscriber Compensation: Recognizing the frustration and wasted tokens, Anthropic has reset usage limits for all subscribers as of April 23.
“We never intentionally degrade our models,” Anthropic stated, emphasizing that the goal is to ensure the user experience matches the high standards of their underlying technology.
Conclusion: Anthropic has identified that recent performance drops were caused by optimization attempts in the software layer rather than the AI itself. By reverting these changes and tightening internal testing, the company aims to restore Claude’s reputation for high-level reasoning.
