Anthropic Confirms Claude Degradation Was Triggered by Harness and Prompt Changes, Not Model Weights

Published on 24 April 2026

Three “small” product layer tweaks quietly altered reasoning

Developers reported “AI shrinkflation” as Claude appeared less capable, more repetitive, and less efficient with tokens. Anthropic’s technical post-mortem says the model weights didn’t regress, but three surrounding product-layer changes did: a reasoning-effort default, a caching bug that wiped thinking too often, and tighter verbosity limits. The company says it has reverted the fixes and reset subscriber usage limits.

Anthropic blames quality drops on product-layer changes, not model weights
A caching logic bug repeatedly cleared “thinking,” harming short-term context
Reasoning effort and verbosity limits reduced performance on complex coding
Fixes include reverting changes and resetting subscriber usage limits

#anthropic #developer tools #ai models #llm evaluation #claude

Read the full story at Venture Beat

This summarization was done by Beige for a story published on Venture Beat

Anthropic Confirms Claude Degradation Was Triggered by Harness and Prompt Changes, Not Model Weights

The full experience is on mobile.