Trending Now - Anthropic Addresses Claude's Quality Issues With Key Fixes
WHO:
Business Insider - Stella Laurenzo, a Senior Director in AMD’s AI group, conducted an audit of 6,852 Claude Code session files and over 234,000 tool calls on GitHub.
VentureBeat - Anthropic, a company specializing in AI models, addressed quality issues with its flagship model, Claude.
VentureBeat - Stella Laurenzo, a Senior Director in AMD’s AI group, conducted an audit of Claude's performance.
WHAT:
VentureBeat - Anthropic confirmed that the underlying model weights of Claude had not regressed, but three specific changes to the product layer caused performance issues.
Business Insider - Anthropic stated that as of April 20th, 2026, the identified issues had been fixed.
VentureBeat - Anthropic published a technical post-mortem identifying three changes that caused quality issues in Claude.
VentureBeat - Claude Opus 4.6's accuracy dropped from 83.3% to 68.3%, causing its ranking to fall from No. 2 to No. 10 according to BridgeMind.
VentureBeat - Anthropic reset usage limits for all subscribers on April 23rd, 2026, to compensate for token waste and performance issues.
WHERE:
VentureBeat - The issues were reported by users across platforms including GitHub, X, and Reddit.
WHEN:
Business Insider - The controversy gained momentum in early April 2026, with user complaints surfacing prominently.
VentureBeat - The controversy began in early April 2026.
VentureBeat - Anthropic changed the default reasoning effort on March 4th, 2026.
VentureBeat - A caching logic bug was introduced on March 26th, 2026.
VentureBeat - System prompt verbosity limits were added on April 16th, 2026.
WHY:
VentureBeat - Anthropic acknowledged that the changes made to Claude's default reasoning effort, caching logic, and verbosity limits inadvertently hampered performance.
VentureBeat - The default reasoning effort was changed to address UI latency issues, which inadvertently reduced Claude's intelligence for complex tasks.
VentureBeat - A caching optimization was intended to improve performance but contained a critical bug that affected Claude's memory.
VentureBeat - The verbosity limits were intended to reduce output length but resulted in a drop in coding quality evaluations.
HOW:
VentureBeat - Anthropic is implementing internal dogfooding, enhanced evaluation suites, and tighter controls to prevent future regressions.
VentureBeat - Anthropic identified three specific changes to the "harness" surrounding Claude that hampered performance: default reasoning effort, a caching logic bug, and system prompt verbosity limits.
VentureBeat - Anthropic plans to implement internal dogfooding, enhanced evaluation suites, and tighter controls to prevent future regressions.
(UPDATED April 24, 2026 09:27 UTC)
A recent audit by Stella Laurenzo, a Senior Director in AMD’s AI group, examined over 6,800 Claude Code session files and more than 234,000 tool calls on GitHub, revealing performance issues linked to three specific changes in the product layer. Anthropic confirmed that while the underlying model weights of Claude had not regressed, these changes inadvertently affected performance, prompting user complaints that gained traction in early on Monday26. As of on Monday, Anthropic has addressed the identified issues and is now implementing internal dogfooding, enhanced evaluation suites, and tighter controls to prevent similar problems in the future.
--
(UPDATED April 23, 2026 23:20 UTC)
Anthropic has acknowledged significant quality issues with its AI model, Claude, which saw its accuracy plummet from 83.3% to 68.3%, dropping its ranking from No. 2 to No. 10 according to BridgeMind. The company published a technical post-mortem detailing three changes that contributed to these problems, including a caching logic bug and new verbosity limits, and has reset usage limits for all subscribers to address the situation. Senior Director Stella Laurenzo conducted an audit of Claude's performance, and Anthropic plans to implement measures like internal dogfooding and enhanced evaluation suites to prevent future regressions.