Inconsistent Usage Reporting Across Audio Transcription Models (Zero Tokens Reported for gpt-realtime-whisper)

API
, , ,
1

Hello Dear Team

We are tracking and reporting usage across different audio transcription models coupled with our Realtime workflows. During testing, we noticed a significant inconsistency in how usage metrics are returned in the API response object depending on the specific transcription model selected.

The Issue

  • whisper-1 correctly reports usage calculated in seconds ("usage_type": "duration").

  • gpt-4o-transcribe correctly reports usage calculated in tokens ("usage_type": "tokens").

  • gpt-realtime-whisper returns all zeros for both tokens and duration, despite executing identical calls with successfully populated transcripts.We made 3 calls with each of the transcription models.

Is gpt-realtime-whisper missing usage metrics by design in the current API version, or is this an active telemetry bug?

We have isolated three sequential calls demonstrating this behavior below.

Isolated Test Logs

JSON

[
  {
    "comment": "CALL 1: whisper-1 properly tracks usage via duration_seconds",
    "call_id": "13",
    "usage": {
      "transcription": {
        "usage_type": "duration",
        "total_tokens": 0,
        "input_tokens": 0,
        "output_tokens": 0,
        "duration_seconds": 13,
        "events_count": 2,
        "model": "whisper-1"
      }
    }
  },
  {
    "comment": "CALL 2: gpt-4o-transcribe properly tracks usage via tokens",
    "call_id": "15",
    "usage": {
      "transcription": {
        "usage_type": "tokens",
        "total_tokens": 142,
        "input_tokens": 108,
        "output_tokens": 34,
        "duration_seconds": 0,
        "events_count": 2,
        "model": "gpt-4o-transcribe"
      }
    }
  },
  {
    "comment": "CALL 3: gpt-realtime-whisper returns zeroed metrics for both types",
    "call_id": "17",
    "usage": {
      "transcription": {
        "usage_type": "tokens",
        "total_tokens": 0,
        "input_tokens": 0,
        "output_tokens": 0,
        "duration_seconds": 0,
        "events_count": 2,
        "model": "gpt-realtime-whisper"
      }
    }
  }
]

Question regarding Architectural Best Practices

As a secondary question regarding voice-agent optimization: when wiring transcription components to complement native gpt-realtime workflows, which model is recommended for the absolute highest phonetic and contextual accuracy for streaming voice agents?

Any insights from the engineering team on the usage reporting or model recommendations would be highly appreciated!

Thanks so much

God bless