Anthropic engineer explains why Claude's writing got worse although the model got smarter
Claude got smarter at math and code. It got worse at writing. Here's why — and what Anthropic is doing about it.
Why it matters
A frontier lab engineer publicly explains the capability-quality tradeoff in recent model versions: optimizing for technical reasoning created a style that alienates human users. This shapes how practitioners evaluate which Claude version to deploy and reveals Anthropic's training choices.
The key facts
10 to knowJackson Kernion (Anthropic engineer) attributes writing degradation to optimization for math, code, and technical explanations
Newer Claude models produce 'overly-dense info dumps' optimized for AI-to-AI communication, not human readability
Opus 4.6 identified as superior pure writing model compared to later versions
Opus 5.5 attempts to address the writing quality regression
Tradeoff between technical capability and prose quality in model training/alignment
Jackson Kernion (Anthropic engineer) cites optimization target shift: newer Claude models tuned for math, code, and technical explanations aimed at other AI models
Result: human-facing writing described as 'overly-dense info dumps'
Opus 4.6 remains the benchmark for pure writing capability
Opus 5.5 attempts to rebalance the tradeoff
Tradeoff is architectural/training-driven, not a bug — reveals frontier lab design choices
Go to the source
The Decoderthe-decoder.com
Publisher excerpt: Anthropic employee Jackson Kernion explains why newer Claude models write so oddly. Optimizing for math, code, and technical explanations aimed at other AI models has created a style that sounds like "overly-dense info dumps" to humans. Opus 5.5 tries to fix this, but Opus 4.6 remains unmatched as…