Anthropic details unannounced AI changes that dumbed down answers
Updated
Updated · Computerworld · May 4
Anthropic details unannounced AI changes that dumbed down answers
4 articles · Updated · Computerworld · May 4
An April 23 report said Claude Code’s default reasoning was cut from high to medium on March 4 and restored on April 7 after users complained about weaker coding responses.
Anthropic also said a March 26 memory-clearing bug made Claude forgetful and repetitive until April 10, while an April 16 prompt tweak reducing verbosity hurt coding quality and was reversed on April 20.
The company said the changes aimed to reduce latency, token usage and frozen interfaces, but highlighted how vendors can alter paid AI systems without advance notice, complicating enterprise oversight, reproducibility and cost control.
As AI token costs skyrocket and model behaviors shift, how can enterprises regain control and ensure transparency over mission-critical systems?
Could sudden, unannounced AI model changes by vendors put your business at legal or financial risk without warning?
Is the current token-based pricing model for enterprise AI creating hidden costs and incentives that undermine long-term trust and stability?
How Anthropic’s March-April 2026 Updates Caused a 67% Decline in Claude Code Performance
Overview
In early 2026, Anthropic faced severe GPU shortages that increased costs, prompting a reduction in Claude's default reasoning effort in March, which led to shallower outputs. A caching bug later caused Claude to lose context mid-session, draining user limits faster. In April, to address latency, Anthropic imposed strict verbosity limits that fragmented Claude's responses and disrupted coding workflows, causing a 3% quality drop. These combined issues resulted in a 67% decline in Claude's thinking depth, increased errors, and a sharp fall in benchmark accuracy. User frustration surged, leading to backlash and threats of cancellations. Anthropic responded by rolling back changes, patching bugs, resetting usage limits, and committing to greater transparency and improved update governance to restore trust and performance.