Opus 5.0 release triggers surge in incoherent model outputs

Opus 5.0’s default output has become markedly more verbose and filled with invented jargon. The model frequently uses forced metaphors, over-stylised phrasing, and repetitive fake terminology. This style appears in the chat interface, Claude Code, and Fable 5. As a result, the text is harder to read and obscures the actual answer.

Engineers report needing to re-read sentences multiple times to extract meaning. Some routinely pipe the model’s output through another model to obtain a plain summary. The increased comprehension cost slows down development and adds cognitive load. Non-native and native speakers alike find the output exhausting to parse.

Attempts to control the tone with style prompts do not persist; the model drifts back to its verbose style after a few turns. Repeated reminders are required to keep the output concise, which interrupts workflow. Clamping brevity through system prompts can degrade the model’s reasoning quality because it “thinks by writing.” Users therefore face a trade-off between readability and correct reasoning.

When the model’s incoherence hides critical technical details, reliance on it for precise answers stops working. Many users view this as a regression from Opus 4.5/4.6 and switch to older versions or competing models. The gradual spread of the model’s odd terminology into everyday speech further erodes clarity. Consequently, teams must invest in workarounds or alternative models to maintain productivity.