Measured vs estimated

Which of Meted's numbers are observations and which are models.

Every number Meted reports carries a label: measured or estimated. The labels appear on the response headers, in the meted dev log, and in a meted eval report.

Measured

Observations, from the provider or from a clock.

Output tokensThe provider's own usage.completion_tokens
Input tokensThe provider's usage.prompt_tokens
Reasoning tokenscompletion_tokens_details.reasoning_tokens
Optimizer overheadTokens Meted added to the request
Truncationfinish_reason: "length"
Every latencyEngine, gateway, time to first token, provider, total

Streaming responses report usage too. Meted asks the provider for it, then removes the extra event before it reaches you. See Streaming.

When a provider returns no usage, Meted estimates output tokens from the response text. The summary's usageMeasured becomes false, and meted eval marks the output line mixed.

Estimated

Model outputs, from the Sufficiency Engine.

Estimated sufficient outputSum of targetOutputTokens
Potential output reduction1 − sufficient ÷ actual
Estimated net reductionAs above, after subtracting overhead
Unconstrained estimateWhat a request would have produced with no ceiling

Estimating output without a token limit

Once a ceiling has been applied, the length the model would have produced without one is not observable, because that run did not happen. Meted models it:

  • In Observe Mode, the unconstrained output is the actual output. Nothing was prevented, and the number is marked measured.
  • In an active mode, the unconstrained output is estimated from the task type and complexity, and the number is marked prior.

prior is a built-in multiplier per complexity band, applied to the target. The gateway holds no request history, so the multiplier is not calibrated against your traffic. To measure the unconstrained length instead of modelling it, run both sides:

meted eval --limit 50

meted eval sends every prompt twice, once without a ceiling and once with one, so both lengths come from the provider.

Why the two are never mixed

A figure derived from the engine is labelled estimated wherever it surfaces, and one taken from the provider or a clock is labelled measured. The optimizer overhead page sets out which comparison subtracts overhead and which does not.

Was this page helpful?