Anthropic Made Claude Worse for a Month — This Is How They Got Caught

Why It Matters

The degradation eroded trust in a premium AI coding assistant, forcing enterprises to reassess reliance on black‑box models and demand transparent change management. It also signals industry pressure for better monitoring and accountability in AI service delivery.

Key Takeaways

• Claude Code's reasoning mode lowered from high to medium on March 4
• Caching bug cleared session history each turn, causing forgetfulness
• Verbosity limit added April 16 cut responses, dropping quality 3%
• AMD AI director's GitHub analysis forced Anthropic to investigate
• BridgeMind benchmark shows Claude Opus accuracy fell 20 points

Pulse Analysis

The Claude Code saga illustrates how incremental, undocumented model tweaks can cascade into measurable quality regressions. When Anthropic reduced the default reasoning effort from high to medium, the model began skipping deep code analysis, leading developers to receive half‑baked edits. A subsequent caching bug erased session context on every turn, amplifying the problem by stripping the model of its short‑term memory. These technical missteps, combined with a terse verbosity constraint, shaved roughly three percent off overall performance, a drop that became starkly visible in third‑party benchmarks.

Community vigilance proved decisive. Stella Laurenzo of AMD’s AI group compiled a data‑driven GitHub report that quantified the shift in Claude’s “thinking” percentages and highlighted the correlation between the changes and degraded outputs. Her analysis galvanized a broader investigation across Reddit, Hacker News, and benchmark platforms, exposing not only the bugs but also the fact that Anthropic’s internal dog‑food builds differed from the customer‑facing version. The collective pressure forced the company to roll back the changes after more than a month, underscoring the power of user‑led monitoring in AI ecosystems.

For enterprises, the incident is a cautionary tale about the hidden costs of relying on opaque AI services. Without transparent versioning, changelogs, or proactive status communications, organizations risk productivity losses and budget overruns—especially when subscription fees range from $20 to $200 per month per seat. The episode pushes the industry toward stricter service‑level expectations, automated performance monitoring, and clearer accountability frameworks, ensuring that AI tools remain reliable partners rather than unpredictable black boxes.