@icy-policeman-64898 how to get the gemini live api's tocken usage for audio input and audio output with the logs

Thanks for reaching out, Abhishek. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.

for gemini-3.1-flash-live-preview

Great — thanks for specifying

gemini-3.1-flash-live-preview

. Here’s how to get audio input/output token usage and related logs for that model. What you’ll use • The

metrics_collected

event emitted by an

AgentSession

. It contains a

RealtimeModelMetrics

payload with per-modality token stats (audio input and audio output). • The Agent Console’s Usage pane and Agent observability to retain and view session metrics after the session ends. Steps to retrieve token usage and logs 1. Ensure Agent observability is enabled for your project in the Agent Console so session metrics and logs are retained. 2. While a session is running, listen for

metrics_collected

events from the

AgentSession

. Parse the

RealtimeModelMetrics

payload to read token usage fields for: ◦ audio input tokens ◦ audio output tokens 1. For historical or aggregated views, open the Agent Console → Usage pane to see cumulative token/audio usage for the project and model (

gemini-3.1-flash-live-preview

). 2. If you need per-session transcripts or detailed logs, ensure observability is on; those artifacts are not retained by default after session end unless observability is enabled. If you want, I can: • Show an example

metrics_collected

payload and the exact fields to parse. • Walk through enabling Agent observability in your console if you tell me whether it’s already enabled. Which would you like?, with interactive elements

Show an example

Below is an example

metrics_collected

event payload (JSON) you would receive from an

AgentSession

that includes

RealtimeModelMetrics

with per-modality token stats for

gemini-3.1-flash-live-preview

. Parse the

audio_input_tokens

and

audio_output_tokens

fields (names shown below) to get token usage for audio.

{
  "event": "metrics_collected",
  "timestamp": "2026-05-10T14:23:45.123Z",
  "session_id": "agent-session-abc123",
  "agent_id": "agent-xyz789",
  "model": "gemini-3.1-flash-live-preview",
  "realtime_model_metrics": {
    "request_id": "req-0001",
    "duration_ms": 1250,
    "latency_ms": 240,
    "total_tokens": 1840,
    "per_modality": {
      "text": {
        "input_tokens": 120,
        "output_tokens": 80
      },
      "audio": {
        "audio_input_tokens": 960,
        "audio_output_tokens": 680
      },
      "vision": {
        "input_tokens": 0,
        "output_tokens": 0
      }
    },
    "billing_breakdown": {
      "text_tokens": 200,
      "audio_tokens": 1640
    }
  },
  "session_state": {
    "is_active": true,
    "participants": 2
  }
}

What to extract • audio input tokens:

realtime_model_metrics.per_modality.audio.audio_input_tokens

• audio output tokens:

realtime_model_metrics.per_modality.audio.audio_output_tokens

• total tokens (if needed):

realtime_model_metrics.total_tokens

• timestamp/session id for correlating logs:

timestamp

,

session_id

If you want, I can: • Show a short parsing snippet in your preferred language (Node, Python, or Go) to extract those fields. • Show how to aggregate these metrics across multiple

metrics_collected

events for a session or project. Which would you prefer?, with interactive elements