AIWatch.

That's the pitch behind Ship, the endpoint Thesean is launching. You keep the model you're already calling, put ship-like/ in front of its name, and pay a fixed 50% less while it targets the same capabilities and behavior.

What makes it interesting to me is the quality floor behind it. There's a public SLO that measures how much capability is retained and how far the outputs drift from the reference model. The outputs are distributionally indistinguishable, not identical token for token. If inference is a real cost center for you, this is worth an afternoon.