Cost Series 04 — Classifiers & Routing

Working notes for a planned episode — an outline, not an article. Details will shift when it gets scripted. Series index: The Bill.

The single biggest lever: send each call to the cheapest model that clears the bar. A classifier is a cost-optimizer wearing a different hat.

The hook

Between a small model and a frontier one is a ~15–20× price gap. If every request goes to the big model "to be safe," you're overpaying by more than caching or output trimming could ever recover. The fix is a classifier that routes — and you probably already run one without calling it that.

What the viewer walks away with

• Model tier is the ~15–20× lever — bigger than every other knob combined
• Routing = classification: decide per call which model/lane clears the quality bar
• Cheap-first-then-escalate: try the small model, escalate only on a failed check — pairs with the evaluator (episode 06)
• Classifiers show up everywhere: routing, gating (security), semantic cache hits — same shape

Beats

• The "everything to the big model" default and what it costs
• The router as a classifier: one endpoint, many models, a routing decision per call
• A security classifier as a worked example: a classifier that changes where work goes for reasons other than cost, same machinery
• Cheap-first-escalate: the pattern, and why it needs a checker to be safe (bridge to 06)

Demo ideas

• The local router auto-routing a mixed workload; cost-per-task by chosen model
• Cheap-first-then-escalate on a task batch: how many clear the small model, what escalation costs

Source material

• The local LLM router — routing/auto-routing as cost control
• A security classifier deciding the lane — same machinery, different objective
• Fleet Series 09 (the router) and 04 (ensembles) — adjacent machinery

Open questions

• How to frame quality bar / escalation criteria without a full evals detour (save depth for 06)
• Include semantic caching (classifier decides a cache hit) here or as its own beat