Reasoning reuse: stop paying for the same chain of thought twice
Chain-of-thought costs several times more tokens than the answer it produces, and it's usually thrown away. For structurally similar problems, most of that reasoning can be reused.
Reasoning is the most expensive thing a model does. A chain of thought can burn several times more tokens than the final answer, and then it's discarded. The next structurally identical problem starts from a blank page and pays the whole cost again, even though the shape of the solution is the same, only the numbers changed.
Reuse the shape of the reasoning; pay only for the last mile.
It's the difference between re-deriving a method from scratch and applying a method you already worked out. For workloads full of structurally repetitive problems, the same kind of calculation, lookup, or multi-step task with different inputs, reusing the reasoning is where the deepest savings hide, well beyond simple answer caching.
The bottom line
You already paid to figure out how to solve this shape of problem. Reuse that work, and let the model spend tokens only on what's genuinely new.