The Model Portfolio: How to Route, Mix, and Govern LLMs Like a Strategic Asset Allocator

A new flagship ships. By Friday you have migrated everything to it.Then latency creeps. The bill climbs on calls that never needed the smartest model. A capability you relied on regresses in an update you cannot refuse. You are running production on one vendor's roadmap and calling it simplicity. It is exposure. It quotes no savings figure from anyone else's benchmark; it gives you the method for estimating your own. The standard move is to pick the best model and standardize. That fails, because there is no best model. There is only one best for a task, a budget, and a failure tolerance, and they stop agreeing once traffic is real.Allocate models the way an investor allocates capital The Model Portfolio treats your models as assets allocated on purpose, not one bet you keep doubling. It introduces ROUTE, five connected practices that turn a stack of API keys into a managed system. - Register. Give every model a mandate, a cost class, a privacy class, and a retirement date, so nothing runs unowned.- Objective typing. Type each task by what it demands: reasoning depth, output format, latency, quality threshold, privacy class. Not the loudest launch.- Utilize policy. Write a routing policy that sends work to the cheapest model clearing the bar, with cascades that escalate only when work earns it and fallbacks that hold when one fails.- Track. Meter cost, latency, and quality per route, so a waste pattern surfaces before the invoice does.- Evolve. Review the mix on a quarterly cadence with promote, demote, and retire criteria, so a deprecation notice becomes a scheduled release instead of a scramble. You finish with a model portfolio whose routes can be explained, tested, and changed. When the leaderboard turns over, your architecture holds.What this book does not doIt promises no savings from someone else's benchmark: cascade trade-offs are estimated from your own rate card, request split, and quality tests. Jurisdiction-specific legal interpretation stays outside the book's scope and belongs with your own counsel. And on low volume, one task type, and no regulated data, the overhead may genuinely not pay. The book says so, and says where that case ends.Who it is forFounders, technical leads, and AI architects who run more than one model in production. If you do, something already routes the work. The only question is whether anyone chose it. The router is infrastructure; the routing policy is governance. From the LPH98.ventures imprint - Systems for the Strategic Self