"Which AI Model Is Best?" Is the Wrong Question

"Which AI model is best?" is the wrong question. The better one: what shape is this task, and which tier of model fits it?

Call it capability-tier routing. You sort models by cost and capability, then match each task to a tier by how much a mistake there would cost. Not by a leaderboard. The model becomes a config setting. The routing is the strategy.

Here's the map I run:

Frontier tier: compliance checks, final fact verification, judgment calls. Pay the premium only where a mistake costs more than the tokens. Nowhere else.

Fast and cheap tier: extraction, classification, formatting. High volume, clear spec. A frontier model here just inflates the bill.

Mid-tier: long drafting and long-context work, where holding coherence matters more than brilliance. Capable enough, cheap enough.

Judge tier, a different model family than the one that wrote the draft: the critique pass. Call it the echo rule. A second opinion from the same model shares its blind spots. You want a different voice, not a louder version of the same one. It costs double the calls, so use it only where errors are expensive.

Open-weight tier: bulk, cost-sensitive jobs.

I built a content system on exactly this logic. It routes about 15 models from roughly 8 providers through a single gateway on one key, sorted into 5 tiers. Swapping a model is editing a config map, not a rewrite. Every call logs its cost, tokens, and latency to a database, tagged by model and tier. A judge model from a different family handles the compliance pass. The echo rule, applied.

The honest catch: this is overhead. The gateway, the logging, the tiering cost real engineering time. If you have one simple job, pick one good model and be done. This only pays off at volume, when your tasks genuinely differ in shape and consequence.

Does it work? The tiering keeps cost predictable. The logging makes regressions visible. When a better model ships, adopting it takes minutes, not days.

The model is config. Pick the tier. Log the call. Swap when something better ships.