LLM cost optimization in Los Angeles.

The technical reality of media-scale spend in LA.

Questions from Los Angeles buyers.

Multimodal calls are expensive. How do we control vision costs on image and video workloads?

Resolution discipline first, model sizing second, frame strategy third. Vision pricing scales with image tokens, which scale with resolution: most tagging, moderation, and description tasks perform identically on a downscaled image at a fraction of the tokens, and the pipeline should downscale by default with full resolution reserved for tasks that proved they need it. Model sizing mirrors text: small multimodal models handle classification and descriptive tagging at large discounts to frontier vision models, gated by your labeled sample. For video, the unit economics live in frame selection: sampled keyframes plus a transcript usually beat dense frame extraction by an order of magnitude, with dense sampling reserved for the segments that matter (we detect those cheaply first). An archive pass that looked like six figures at naive settings routinely lands at a tenth of that after these three disciplines, and the audit shows the per-asset math before anything runs.

We want to backfill metadata across a large archive. How do we keep the one-time cost sane?

Treat the backfill as a batch-priced, tiered campaign rather than a single pass. The structure: batch APIs take the whole job at roughly half synchronous price, since nothing about a backfill is real-time; a cheap first tier (small model, downscaled assets) generates baseline metadata across everything; and an escalation tier re-processes only the assets the first pass flagged as ambiguous or high-value (a few percent, typically) with stronger models. Add resumability (checkpointed progress, idempotent writes) so failures never re-bill completed work, and run a priced pilot on a representative slice first, because per-asset cost measured on 10,000 assets makes the full-archive budget a multiplication rather than a hope. Archives that penciled out as unaffordable at naive frontier pricing routinely become a defensible line item under this structure, which is the difference between the backfill happening and staying a someday.

What does tagging actually cost at catalog scale, per thousand items?

On a right-sized pipeline: low single-digit dollars per thousand text items, and tens of dollars per thousand images, with naive pipelines running ten to twenty times both. The drivers are unglamorous. Text tagging on a small model with a cached schema prompt costs fractions of a cent per item; the same task on a frontier model with the taxonomy re-explained every call costs multiples of that for identical output. Images add vision tokens, which resolution discipline keeps in check. Batch pricing halves both. The compounding matters at LA volumes: a fashion catalog refreshing 200,000 SKUs monthly, or a media library tagging at ingest, lives with these per-thousand numbers as a recurring bill, so the gap between naive and engineered is not a rounding error; it is headcount-sized money. The audit prices your actual items on your actual taxonomy before recommending anything.

Our brand and style context makes prompts huge. Can caching tame that?

This is caching's favorite case. Brand voice documents, style guides, taxonomy definitions, and few-shot examples are exactly the heavy, stable, reused-on-every-call content that prompt caching discounts hardest: structure them as a byte-identical prefix and the 90-percent input discount applies to the bulk of your tokens, with the per-item content (the SKU, the asset, the copy brief) riding cheap in the volatile suffix. The engineering is consolidation and stability: one canonical brand-context block per brand (not per-team variants that fragment the cache), versioned deliberately (a changed comma is a cache miss across all traffic), and ordered before any volatile content. Multi-brand shops layer it per brand and still cache excellently within each. Teams with thousand-token brand prefixes typically watch input costs drop by more than half from this restructuring alone, with zero change to output quality, because the model sees identical content either way.