inference.Response carries the same usage pointer both at the response level and, when present, on its content.AIMessage. The bundled decoders clone the normalized value for the message so callers do not share mutable pointer state.

Shape

type Response struct {
	Message      *content.AIMessage
	Usage        *content.Usage
	Model        string
	FinishReason stream.FinishReason
	Attempts     int
}

Usage == nil means the provider did not report usage. It does not mean zero tokens. A malformed or unrepresentable usage payload — a null where a count is required, a fraction, a negative, an out-of-range value, or a subset larger than its gross total — returns a typed normalization error instead of a partial response.

Normalization

OpenAI Chat subtracts cached and cache-write subsets from gross prompt tokens. Responses subtracts both cached input tokens and cache-write tokens. Anthropic and Bedrock report cache subsets as separate fields. Gemini subtracts cached content, adds the separately reported tool-use prompt tokens, and adds candidate and thought output counts; its reported total is validated as a well-formed count but is deliberately not reconciled against those components. No decode path gates on the reasoning-within-output convention: a provider that reports more reasoning tokens than output tokens has an accounting bug, not an invalid response, so both counts are carried as reported. Test the convention with content.Usage.ReasoningWithinOutput().

if response.Usage == nil {
	// Provider did not report usage; retain nil in telemetry.
	return nil
}
contextTokens, err := response.Usage.ContextTokens()
if err != nil {
	return err
}
fmt.Println(contextTokens)

Source and proof

Run go test ./codec/....