I think what’s pretty cool about Ship is that it’s finding the least expensive way to produce behavior that is effectively equivalent to frontier models. Not every request is worth the same number of tokens: the value and difficulty of inputs vary, and many requests simply do not need frontier-level reasoning.

At inference time, Ship can choose how much compute a request actually deserves and which method should handle it—a model, a tool-using harness, a cascade, an ensemble, a program, or some combination.

It reminds me of how we thought about introducing reasoning models with o1. Most chat experiences did not need o1; only the questions complex enough to justify the extra reasoning.

The natural business model is dynamic too: price the intelligence actually used, not every request as though it required the most expensive possible system.