AI Explained: Claude Opus 4.7 'Shrinkflation' — SWE-bench 87.6% but Creative Writing and Web Research Regressed

AI Explained·medium signal

AI Explained's deep analysis documents specific regressions in Claude Opus 4.7: creative writing lost warmth and defaults to bullet points, web research source attribution declined, and contradiction detection weakened. Benchmarks tell a different story — SWE-bench Verified hit 87.6% (up from 80.8%), GPQA Diamond 94.2%, and CursorBench 70%. The tradeoff is deliberate: Anthropic optimized for coding/agentic performance at the expense of prose and research quality.