LLM cost optimization in Portland.

The technical reality of lean-brand spend in Portland.

Questions from Portland buyers.

What is the cheapest model we can actually rely on, and how do we find it for our work?

It is task-specific, and the finding method matters more than the answer. The method: take two hundred real examples of your task (product copy, spec extraction, support drafts), have the person who owns quality grade outputs from three or four model tiers blind, and let the grades pick the floor. The pattern that emerges for most Portland content and document work: current small models (Haiku-class and peers) clear the quality bar for templated, grounded, or extraction-shaped tasks at one-tenth to one-thirtieth of frontier pricing, while open-ended brand writing and tricky judgment calls still earn the mid or frontier tier. The result is a two-tier menu rather than a single loyalty: the cheap reliable workhorse on the volume, the strong model on the flagged minority. We run this benchmark as the audit's first week, and the blind grading regularly retires expensive assumptions, in both directions.

Our brand guide and tone examples ride along on every generation call. What does caching that save?

Usually more than half the input bill, because brand context is the heaviest and most stable thing in a content pipeline. The mechanics: prompt caching discounts identical prefixes by 90 percent on current provider pricing, and a brand guide plus graded examples plus taxonomy is exactly the content that never changes between calls. The engineering: consolidate to one canonical brand block (not the five drifting variants different teammates paste), order it before any volatile content (the SKU, the brief), keep it byte-identical (a tweaked comma is a cache miss across all traffic), and version it deliberately when the brand evolves. For multi-brand houses, each brand gets its own cached block and the discount applies within each. Teams running thousand-token brand prefixes on every call typically watch input costs drop 50 to 70 percent from this restructuring alone, with output quality untouched, since the model reads identical content either way.

Product copy refreshes for drops and seasons: how cheap can a batch run get?

Cheap enough that the bottleneck becomes review, not generation. The engineered stack for a Portland-scale catalog refresh: a small model gated by your merchandiser's blind grading, the brand and taxonomy context cached at the discounted rate, and the whole run submitted through batch endpoints at half price. Arithmetic for a 20,000-SKU seasonal refresh: titles, attributes, and descriptions land in the tens of dollars of inference, processed overnight with checkpointing so a failure never re-bills finished SKUs. The honest cost center that remains is the human one: sampling review for the long tail and new categories, which we tune by risk so creative hours concentrate where doubt lives. The practical effect: refreshes that used to be quarterly negotiations with an agency budget become a routine pipeline run, and the brand voice survives because it is engineered into the cached context rather than re-improvised per batch.

More than savings, we want no surprises. Can a small team get a genuinely predictable AI bill?

Yes, with four lightweight controls that take a day to install and run unattended. Pace alarms: provider budgets with alerts at 50 and 80 percent of the monthly number, routed to a human inbox, so drift is a Tuesday email rather than a month-end shock. Per-workflow visibility: spend split by pipeline (product copy, support drafts, spec extraction) in a simple dashboard, so growth is attributable the day it starts. Hard ceilings: per-job token caps and retry budgets, because the classic small-team bill spike is a retry loop or an unbounded batch, not gradual growth. And change discipline: any prompt or model change ships with a before-and-after cost-per-item note, since most mystery increases are a deploy wearing a costume. With those four in place, the monthly bill becomes a forecast with a small error bar, which for a lean Portland team is frequently worth more than the optimization itself, and we install them as standard in every engagement here.