Models
Distillation
Distillation trains a smaller model to imitate a larger one, producing something much cheaper to run that retains most of the capability on a narrower range of tasks.
It is how the cheap tiers of most model families exist. The small model learns from the large model’s outputs rather than from raw data, which transfers behaviour far more efficiently than training from scratch.
The capability gap is real but narrower than the price gap, which is what makes model routing worthwhile: send the easy majority to the distilled model and reserve the flagship for what needs it.
In practice
Distillation is where the cheap tier comes from — a small model trained on a large one's outputs, retaining most of the capability at a fraction of the serving cost. It is also why the price gap between tiers keeps widening while the quality gap on ordinary tasks narrows, and why routing easy traffic downward keeps getting more attractive.
Common questions
Are distilled models worse?
On ordinary production tasks, often indistinguishable. On open-ended reasoning and long-context synthesis, usually behind. That gap is exactly the boundary a routing strategy should be drawn along.
Why are cheap models getting so good?
Distillation transfers capability from larger models at a fraction of the serving cost, so the quality gap on common tasks narrows while the price gap widens. That trend is what makes model routing increasingly worthwhile.