OptionalInputPer-modality breakdown of the input tokens, when the provider reports one. Realtime models bill audio and text tokens at very different rates (e.g. GPT Realtime 2.1: $32/M audio in vs $4/M text in), so cost attribution REQUIRES this split — the totals alone force a wrong blended rate. Absent when the provider reports only totals.
Number of input tokens reported in this usage update.
OptionalOutputPer-modality breakdown of the output tokens, when the provider reports one. See RealtimeUsage.InputTokenDetails.
Number of output tokens reported in this usage update.
Incremental usage/telemetry reported during a realtime session.
This is the realtime-specific counterpart to Core's request/response
ModelUsage. Because a realtime session is long-lived and usage is reported in increments (and checkpointed incrementally), this type carries the token deltas a provider emits over the life of the session rather than a single final tally.