The gap has narrowed
Open-weight models now handle the majority of real production tasks well: extraction, classification, summarization, retrieval-grounded answers, and many agent steps. The frontier still leads on the hardest reasoning and on raw breadth, but the distance that matters for most workloads has shrunk.
That changes the question from which model is best to which model is right for this task, at this cost, under these constraints.
When to reach for each
Frontier APIs are the right call when you need peak capability, when you are still finding the use case, or when speed to first value beats everything. They cost more per call and your data leaves your walls, but they get you moving.
Open models earn their place when volume is high, when data cannot leave your environment, when latency and cost must be controlled, or when you want to own the model as defensible IP. Hosted yourself, they can be dramatically cheaper at scale.
The pragmatic answer is both
Mature stacks route. A small fine-tuned open model handles the bulk of high-volume work, and a frontier model is called only for the steps that need it. The result is lower cost, more control, and a quality bar held where it counts. The skill is in the routing and the evaluations, not in picking a single winner.
