Why every figure says estimated

Every saving in LightHouse carries the word estimated, in the interface and in the API (estimated: true). None is ever on an invoice or in a billable total (counted: false). That is not modesty. A saving is the difference between what happened and something that did not happen, and the thing that did not happen cannot be measured.

What makes each figure an estimate

Source of uncertainty	Where it applies	Which way it errs
The counterfactual was never run. Nobody sent the stuffed request, ran the top model, skipped the cache, rented the hosted runtime or kept the services always on	Every lever	The rules are chosen to understate: see each lever’s conservative choices on How a saving is calculated
Prices come from your price card, not the provider’s invoice	Every lever	Neither way: both sides use the same card
A call’s cost may be worked out rather than reported. Where a provider returns tokens but no price, the gateway prices the call from the card; where it returns neither, the call is recorded at the amount reserved for it, an upper bound. Both are flagged estimated	Spend, and every lever built on it	Depends on the card’s accuracy; reconciliation shows the difference
Router savings are projected per call, from the model’s measured tokens on the benchmark suite, not from the call’s own tokens on the model that was not chosen	Model routing (decisions)	Depends on how closely your calls resemble the suite; your own benchmark tasks narrow it
Retrieval is sampled where a deployment samples its search records (sample_rate), and scaled back up	Retrieval	Neither way on average
Corpus size is measured from characters, less chunk overlap, then converted to tokens	Retrieval	Low, by construction
List prices, before credits and discounts	Ephemeral infrastructure	Low: lines common to both profiles are left out of the baseline
Currency conversion at the European Central Bank’s daily rate, for display	Any figure shown in a currency other than dollars	Neither way

What would make it exact

Some figures can get closer to exact; the counterfactuals never can.

• Spend is exact where the provider’s own bill is available. Spend → Reconciliation compares the gateway’s figure with AWS Cost Explorer or Anthropic’s cost report, per day and connection, and shows any difference rather than smoothing it. An invoice is never based on your software’s report alone.
• Routing gets closer when the router’s evidence is your own. Run the Token Value Benchmark with your own tasks, and the router’s cost projections are made on your work rather than the standard suite.
• Routing, measured on your own agents: replaying an agent’s recorded work on a candidate model, with its recorded tool results, and scoring it with the agent’s own checks. That would measure a change of model on that agent directly, rather than projecting it from a suite. It is not built yet: see How a model earns its scores.
• Infrastructure becomes a full comparison, and counted, once you set a comparison card of your own for the hosted platform you would otherwise have used.
• Retrieval cannot be made exact without sending the stuffed request, which is the cost it exists to avoid.

How to talk about it

Say “an estimated saving of $X against [the baseline]”, and name the baseline. That is what LightHouse’s own sentence does, and it is the form that survives a finance review.