When we benchmarked DeepSeek V4 Pro against Claude Haiku 4.5 for AI-first content, the quality gap forced a writer swap. See the scoring framework and roll-out