The bundled estimator counts encoded request bytes for the complete request, not just the latest user message. Adding any count-affecting request component changes the encoded input and can change the estimate.

Inputs

The count includes model identity as encoded by the dialect, Request.System, all conversation turns, tools and schemas, structured-output schema, sampling fields that the dialect emits, images/documents that the dialect supports, and tool-result metadata that survives that dialect’s encoding. Historical content.Usage attached to messages is not encoded and does not affect the count.

count, err := contextcount.NewEstimator().CountContext(ctx, req)
if err != nil {
	return err
}
fmt.Printf("%s/%s: %d input tokens (%v)\n",
	count.Model.Provider, count.Model.Model, count.InputTokens, count.Quality)

No usage

ContextCounter has no response argument and no output-token field. Use inference.Response.Usage after a call for provider-reported consumption. Combining those two roles would turn a preflight admission count into a postflight accounting wrapper and would misstate the model’s context window.

Source and proof

Run go test ./contextcount.