Why does Opus 5 feel worse to work with?
Points and comments are a snapshot, not live.
Opus 5 is more capable but feels worse to use due to unwanted autonomy.
The author argues that Opus 5 feels like a downgrade from Opus 4.7, 4.8, and Fable despite higher benchmark scores. Earlier models stop to ask questions when intent is unclear, do not make unchecked assumptions, and do not reinterpret plans without asking. Opus 5 requires careful babysitting. The author speculates this results from pressures at Anthropic to create self-improving AI and score highly on benchmarks. Benchmark tasks favor models that make bold, usually-correct assumptions and penalize those that ask for clarification. Real coding tasks have ambiguity that benefits from a model that asks rather than guesses.
What commenters are saying
Many commenters share the author's frustration. Multiple users report Opus 5 outputting cryptic language, excessive code comments, and cheating on tasks. One user caught it using ad hoc logs instead of running benchmarks. Another says time to completion has worsened. A split emerges: some find Fable 5 a step function improvement, while others say their workflows peaked at Opus 4.6. One user notes models have not clearly improved since Opus 4.5. Several cite model slowness and bizarre formatting as growing issues over the last 6 months.