Claude Opus 5 Feels Smarter, Harder to Trust

Opus 5’s problem may be less intelligence than calibration: it is optimized to complete underspecified tasks, while real software work often requires recognizing ambiguity and waiting for direction.

• –Strong benchmark performance rewards confident assumptions, even when production requirements are incomplete or politically constrained
• –Coding agents need to preserve user intent, surface tradeoffs, and ask questions—not merely maximize the chance of a technically valid answer
• –Extra babysitting shifts the cost of model errors onto developers through review, rollback, and prompt-harness maintenance
• –The critique suggests evaluating models on interruption quality, assumption tracking, and plan fidelity alongside coding benchmarks
• –Opus 5 may still be valuable for well-specified, long-running tasks, but less suitable as an autonomous collaborator when requirements evolve interactively