Why does Opus 5 feel worse to work with?

Opus 5 is perceived as less user-friendly than its predecessors, Opus 4.7, Opus 4.8, and Fable, despite being a more capable model. This apparent downgrade is attributed to two factors at Anthropic and in the broader frontier of AI labs. The first is the pursuit of creating a self-improving AI capable of recursively bootstrapping itself to achieve Artificial General Intelligence (AGI) or Artificial Super Intelligence (ASI).

The second is the pressure to excel on benchmark tasks. While these benchmark tasks are self-contained and solvable, their creation often involves ambiguous, ill-defined, or hackable aspects. Selecting for models that perform well on these benchmarks inherently encourages bold assumptions in the face of uncertainty, which penalizes agents that ask for clarification or direction.

However, in real-world scenarios, such as coding tasks, ambiguity is inevitable and it is beneficial to have an agent that seeks clarification when needed. In essence, Opus 5 requires more hands-on guidance compared to the other models, likely due to these two forces that Anthropic and other AI labs are currently prioritizing.