01 · Flagship · Penelope router
Route every task to the model that should solve it.
Penelope selects one model and reasoning effort for every task. It needs no training: at inference time, it reads recorded evaluation evidence from similar work and makes one routing decision.
The goal is straightforward: preserve the quality that matters while avoiding expensive inference when a cheaper configuration is enough. Held-out masks, anonymised models, and explicit router cost keep the comparison honest.
Talk to us about routing$40.13Model spend · lowest policy above 65 macro
38–55%Cheaper effort routing at matched accuracy
558kHeld-out evaluation trials in the evidence catalog
58 / 28Models / benchmarks represented
Honest total cost: the router currently adds $32.68. Penelope has the lowest execution spend, but not yet the lowest all-in spend.