Released 15 September 2026, and interesting for what it refuses to do. A question is declared in code rather than written into a prompt string, so the return type is known at compile time and there is no parsing step between the model and the branch that acts on it. TypeSafe reports very large speed and cost advantages over general models on classification — claims worth verifying rather than repeating.
The shape rhymes with something already true here: Zod exists because a value crossing a boundary should be validated at the boundary, and content schemas exist because a contract beats a convention. A model that returns a type rather than a sentence is the same argument aimed at inference. Filed as a seed — a week old at the time of writing, and unproven here.