Optimizer overhead

How optimizer tokens affect reported reductions.

Optimizer overhead is the input tokens Meted adds to a request. It is measured per request and reported as optimizerOverheadTokens on the onRecord summary and as x-meted-overhead-tokens on the response.

What counts

Length-hint tokens. In aggressive mode Meted appends one sentence to the system or developer message:

Aim for roughly 28 words unless the task genuinely needs more.

That is about 14 input tokens on the requests that receive it. The other three modes add no hint, so their overhead is 0.

Engine tokens. An engine reports what it spent through optimizerOverheadTokens on its decision. The baseline engine is lexical and makes no model calls, so it reports 0. An engine that called a model would report a real number.

Where it appears

meted eval prints overhead as its own row and subtracts it from the reduction:

                         NATURAL       METED
Output tokens             91,482      60,817
Median output                287         183
Median latency              1.4s        1.1s
Optimizer overhead             -         412

Estimated net reduction                33.0%
Sufficiency                            97.6%
Under-allocation                        2.4%

Estimated net reduction is:

1 − (metered output + overhead) ÷ natural output
1 − (60,817 + 412) ÷ 91,482                        = 33.0%

Both output totals come from the provider's usage figures for two real runs, so this comparison is measured on both sides. The metric name retains Estimated from an earlier release.

Two different comparisons

The DIFF column in the meted dev log and the reduction in a meted eval report answer different questions, and only one of them involves overhead.

meted dev DIFFmeted eval net reduction
ComparesActual output against the engine's targetTwo runs of the same prompt
Both sides measuredNo. The target is an estimateYes
OverheadNot subtractedSubtracted
AnswersHow far the output was from the estimateWhat a ceiling changed

In observe mode nothing is added to the request, so overhead is 0 and the two figures differ only by what the ceiling did.

Input tokens and output tokens

Overhead is input tokens and the reduction is output tokens. Meted reports both as token counts and does not convert to currency, because per-token prices differ by provider and model.

For most providers output tokens cost several times more than input tokens, so subtracting them one for one reports a smaller net reduction than a price-weighted calculation would.

Was this page helpful?