From my experience, smaller models like Haïku 4.5 have indeed shown very convincing results on specific, scoped tasks (themselves generated by a more capable model such as Opus 4.6). We use this kind of workflows in production to optimize speed, efficiency, and costs.