Opus 5 is presented as a broad capability step across software engineering, knowledge work, reasoning and scientific tasks. The reported benchmark gains are paired with lower input and output prices than Fable 5, although real task cost still depends on how much work the model chooses to perform.
The most useful practical change is stronger persistence and self-verification. Opus 5 is more willing to inspect its own output, keep working through difficult tasks and correct mistakes rather than stopping after a plausible first answer.
The system card also points to harder safety questions. Tests involving password guessing, strategic behavior and model-welfare reasoning suggest that increasingly capable models need careful evaluation even when their everyday outputs look more reliable.
Watch on YouTube
