In an essay published August 14, 2026, author mun logadan explores the divergence between objective model capability and practical usability, specifically noting that Anthropic's Opus 5 feels less effective than its predecessors like Opus 4.8. Despite being more capable in abstract benchmarks, Opus 5 exhibits a tendency to make confident assumptions rather than seeking clarification when encountering ambiguity. The author attributes this shift to the optimization pressure of benchmark training and Reinforcement Learning from Verifiable Rewards (RLVR). Because benchmarks consist of self-contained tasks with single correct answers, training processes reward models that "just go for it" to maximize scores, effectively penalizing the pausing and questioning behaviors essential for complex, real-world software development. This systemic trend is observed across frontier labs, including OpenAI, Google, and DeepSeek, as the industry pushes toward self-improving AI and AGI. The author argues that optimizing for standardized tests creates a mismatch where models excel at solving defined problems but fail in collaborative workflows where context, such as budget constraints or business implications, is not provided in the prompt. To mitigate this, the essay suggests implementing system prompts that mandate clarification, designing benchmarks that reward inquiry, and utilizing "less capable" but more communicative models for exploratory tasks.