A notable shift in how frontier models ship: instead of one model, labs increasingly release a family — a small, a medium, and a large variant sharing a lineage but tuned for different points on the cost-latency-quality curve. The pattern reflects how AI is actually used.
One size never fit all
Most requests hitting a production system are easy. A minority are hard. Serving every request with the largest model is like sending a freight truck to deliver a letter — it works, but the economics are absurd. A family lets you match the model to the request: the small variant for high-volume simple work, the large one reserved for the genuinely difficult.
The frontier isn't a single point anymore. It's a menu, and the skill is ordering well.
Shared lineage, different budgets
Because the variants share training and architecture, they behave consistently — the same prompt tends to work across sizes, so you can develop against one and scale up or down without re-engineering. That consistency is what makes cascading practical: start small, escalate on difficulty, and the handoff doesn't require a different prompt or a different mental model.
What it means for builders
Treat model choice as routing, not a one-time decision. Profile your traffic, send the easy majority to the cheaper tier, and escalate only when a task needs it. The teams getting the most out of frontier models aren't the ones always calling the biggest — they're the ones who built the routing to call the right one.
