Cutting cost per run by 79% without changing a single answer

Bundling every engine into every plan is a good promise and an expensive one. This is how the unit economics were made to survive it, measured across 210 live runs rather than estimated.

The starting position

As first built — a large model, synchronous calls, three samples weekly — a run cost about $0.159. Across six engines at fifty prompts, that is roughly $7,400 per brand per year. Viable at enterprise pricing. Not viable at agency channel pricing, which is the entire go-to-market.

Where the money actually was

Two findings reordered the whole cost model. First, about 70% of the spend was input tokens, and those tokens were *search results*, not our prompts. Second — and this was the useful one — prompt caching could not help at all, because the shared prefix sits below the model's minimum cacheable length. Half the standard cost-optimisation playbook was inapplicable before it started.

Below a certain model size, a third fact takes over: the per-thousand web-search fee becomes the floor, at 42–59% of the bill. You cannot optimise your way past a fee that does not scale with the model.

The two changes

• Submit runs through the provider's batch interface rather than individually. Same answers, roughly half the token cost, at the price of latency — which a weekly cycle does not care about.
• Use a mid-tier model for the analysis pass. Detection, prominence and citation extraction are structured extraction tasks, not reasoning tasks.

Result: $0.0337 per run, about $1,580 per brand-year — a 79% reduction. The check that mattered was whether the output changed: across the bake-off, the mention rate was 30/30 against 30/30. Identical.

One counter-intuitive operational note

Batch latency is not proportional to batch size. In our runs, 18 requests took 12.8 minutes and 150 took 5.0. Small batches are not fast batches, which changes how a cycle should be assembled: fewer, larger submissions beat many small ones.

Why publish this

Because “all engines included, no per-engine add-on” is only credible if the economics behind it are. A bundle whose costs were never modelled is a bundle that gets withdrawn or quietly degraded in month eight, and agencies have seen that film before.