Somewhere on your roadmap sits a line that reads "choose our AI model." Open-weight or closed. Self-hosted or API. Teams burn quarters on the answer. Boards ask about it in the same breath as security. It feels foundational, the way choosing a database felt foundational in 2005.
It is not foundational. It is a routing decision. And routing decisions should stay cheap to change.
The rankings will not hold still.
Closed frontier models lead on hard reasoning today. Open-weight models close the gap on a lag, then win on cost and control for narrower jobs. The leaderboard reorders every few months. An architecture that hard-wires one answer inherits that volatility as risk: you either overpay for routine work or underpower critical work, and re-platforming later costs more than the original decision saved.
The honest answer to "open or closed?" is both, per task. Drafting a positioning brief and tagging ten thousand archived assets are different jobs with different quality bars and different unit economics. A system built on model-agnostic routing sends each job to whichever model clears the bar at the lowest cost, then re-decides when the market moves. Ours runs 13 agents this way: 44,000+ runs in 63 days on under $50 of model spend, because routine work never touches frontier pricing.
The model is the least durable layer.
What persists in a brand system is everything wrapped around the model. Your brand rules, encoded as constraints an agent must satisfy. Your institutional memory: every decision, every correction, every approved output. Your approval gate, so a human signs off before anything ships. Your audit-grade logs, so you can show your work. Those assets survive every model swap. Keep them portable — plain rules, owned data, no lock-in — and swapping models becomes an afternoon, not a migration.
That inverts the buying order. The first question is not "which model?" It is "what does the system remember, and who approves what ships?" Model choice sits downstream of the system that carries your rules and your memory. Get the system right and the model question shrinks to a line item you revisit quarterly.
Two questions that settle it.
If you are evaluating vendors, two questions expose the architecture. First: show me the routing. A serious system can tell you which model handled a given run, why, and what it cost — down to the run. Second: what leaves with us if we leave? The right answer is everything. Rules, memory, logs, connectors. Read-only connectors and exportable logs are not features; they are proof the vendor expects to be judged on results rather than switching costs. The first ninety days should build assets you keep, whatever model runs underneath.
Open versus closed is a real tradeoff — for the engineer picking a model for one task, this quarter. For the operator buying a brand system, it is the wrong first question. Buy the memory, the gate, and the routing. Rent the model.
If your model strategy is currently a procurement debate, reframe it as an architecture question — book a 30-minute strategy blueprint call and we will map it in one session.
