Cost and token optimisation

LightHouse’s Overview opens with one sentence: Saved with Atlantic Agentic Labs: $X and N tokens this month. This section is about how that sentence is made, so that you can check it, explain it to your finance team, and act on it.

Three principles run through every figure:

1. Both sides of every comparison are priced on the same card: your deployment’s own price card. A provider’s price cut moves neither side, so it is never counted as a saving.
2. Every saving is labelled estimated, and none is ever in a billable total. A saving compares what happened with something that did not happen; why every figure says estimated explains what that means and what would make each one exact.
3. The rules err low. Wherever a choice had to be made, it was made so the saving is understated rather than overstated, and each lever lists those choices.

The levers at a glance

Lever	Compared with	In one line	Where in LightHouse
Retrieval	Sending everything you may search with every request	Tokens kept out of the model’s context	Insights → Retrieval
Model routing	Running the most capable model you could have run for everything	The router’s decisions, net of escalations	Agents → Routing
Prompt caching	No cache	The reads’ discount, less the writes’ premium	Spend → Caching
Your infrastructure	A hosted agent runtime’s published rates	Your cloud bill for the stack against theirs	Insights → Savings
Ephemeral infrastructure	The same services always on	List-price estimate less your actual cost	Settings → Deployments

The headline’s dollars are the sum of the levers that have data, signed: a lever that cost more than its baseline lowers the total rather than being left out. The headline’s tokens are retrieval’s alone, because routing saves dollars, not tokens, and cached tokens were still sent.

Start here

What is measuredThe token and cost fields on every model call, and how to group them.

How a saving is calculatedEach lever's formula, baseline and conservative choices.

Reproduce the figureThe API calls, and the arithmetic over your own export.

What to do about itThe levers in order of effect, and where each is set.

How do you know what a token is worth on your own work? That is what the Token Value Benchmark measures.