Martian （@withmartian ） just launched Ship， a drop-in endpoint that replicates your model's output at a 50% lower cost.
It proposes the runtime specialization of inference-time routing.
Most cost tools switch you to a distinctly weaker model to save you money.
Ship doesn't seem to want to make that tradeoff because it goes after the exact abilities and behavior of your reference model.
To get the same grade for cheaper， their system can blend models， tools， harnesses， cascades or ensembles.
Training freezes a model before you ask for it， just like ahead-of-time compilation.
Instead， Ship adds a step at runtime， which is more like a just-in-time compiler changing its focus on the fly.
It all hinges on two guarantees： behavioral equivalence and capability equivalence.