METR AI researcher @ReillyHaskins02 says METR could cut the cost of its live safety monitor by 3x just by routing only suspicious actions to the strongest models: "Our implementation is very basic, and we use frontier adjacent models with the maximum reasoning setting available. We're not really cheaping out on the model or the reasoning." "We could draw down cost a bunch by running faster, cheaper models, or using a lower reasoning effort. You could run the model on every third action…