OpenAI Retires Nine Audio and Realtime Model IDs by January 2027: The Migration Operators Must Complete

On July 20, OpenAI sent deprecation notices to every developer with active traffic on nine legacy audio, realtime, and transcription model IDs. Shutdown date: January 20, 2027. This is the six-month notice OpenAI's policy requires for GA models — which means the window is open now but it will not extend.

The notice landed quietly. The new GPT Transcribe and GPT Live Transcribe models grabbed attention the same week. But the deprecation affects a different and older layer of the stack: the gpt-4o-audio, gpt-4o-realtime, gpt-audio, and gpt-realtime-mini families that many production voice and audio pipelines still call directly.

What gets shut down and when

All nine model IDs retire on the same date:

The operator consequence: what breaks and what doesn't

The endpoint does not change for any of these migrations. Audio-in/audio-out traffic on Chat Completions (v1/chat/completions) stays on the same endpoint — you change the model field only. Realtime WebSocket sessions (v1/realtime) also stay on the same endpoint. This is a model-string swap, not a re-integration.

What does break at shutdown: any request where the model field contains a deprecated string returns an error. If your gateway or application hard-codes gpt-4o-realtime or gpt-4o-audio and you haven't updated by January 20, those calls fail.

The failure mode depends on how your routing handles model errors. A gateway that sends errors back to the client will return a 4xx. A gateway with a fallback policy on model-not-found may silently route to a fallback you didn't intend. Both are problems, but the silent fallback is the harder one to detect — it won't show up as an error spike, only as an unexpected cost or behavior shift.

Three migration paths for three use cases

Path 1: Chat Completions with audio output. Any model value in ["gpt-audio", "gpt-4o-audio", "gpt-audio-mini", "gpt-4o-mini-audio"] maps to gpt-audio-1.5. Pricing on gpt-audio-1.5: $2.50/$10 per 1M text tokens and $32/$64 per 1M audio tokens — same as the current gpt-audio rate card. The context window (128K) and max output (16,384 tokens) are also unchanged. For operators who check pricing-tier entitlements at request time, gpt-audio-1.5 is the current version of the same model; no tier adjustment needed.

# Before
response = client.chat.completions.create(
    model="gpt-4o-audio",   # deprecated Jan 20, 2027
    modalities=["text", "audio"],
    audio={"voice": "alloy", "format": "pcm16"},
    messages=[...]
)

# After
response = client.chat.completions.create(
    model="gpt-audio-1.5",  # replacement
    modalities=["text", "audio"],
    audio={"voice": "alloy", "format": "pcm16"},
    messages=[...]
)

Path 2: Realtime WebSocket sessions. Models gpt-realtime, gpt-4o-realtime, gpt-realtime-mini, and gpt-4o-mini-realtime map to gpt-realtime-2.1 and gpt-realtime-2.1-mini respectively. Pricing for gpt-realtime-2.1: $4/$0.4/$24 text tokens and $32/$0.4/$64 audio tokens — identical to GPT-Realtime-2, which most teams have already been running. The session API contract is the same: connect to v1/realtime, send the session.update event with the new model field.

Path 3: Realtime transcription sessions. The one snapshot-specific migration: gpt-4o-mini-transcribe-2025-03-20 → gpt-4o-mini-transcribe-2025-12-15. This is a minor version bump within the same model family, not a behavioral change. Operators who pinned to the March 2025 snapshot for stability should test the December snapshot before the cutover, particularly for language detection accuracy and turn-boundary behavior, since the December version includes updated VAD calibration.

The router and gateway audit checklist

If you run a multi-model gateway or router, the migration is three steps:

1. Grep your model-ID configs for the deprecated strings. Both the alias names (gpt-4o-realtime, gpt-audio) and any snapshots you may have pinned need to be found. Use the full list in the deprecation table, not just the ones you think you're using.

2. Check your fallback logic for audio-routed paths. A model-not-found error on a realtime session is handled differently by different gateway implementations. Some fall back to a chat completion path, which is not a valid substitute for a WebSocket voice session. Make sure your error handling for realtime paths either hard-fails or falls back to another v1/realtime-capable model — not a chat model.

3. Verify rate-limit entitlements on the replacement IDs. gpt-realtime-2.1 has the same rate limits as GPT-Realtime-2 at each tier (200/400/5000/10000/20000 RPM from Tier 1 to 5). However, if your organization has negotiated custom rate limits on a specific old model ID, confirm those carry over to the new ID with OpenAI before the cutover.

Cross-provider context: how this compares to other realtime audio APIs

OpenAI's approach — six months of notice, same endpoint, model-string-only change — is comparatively clean. Gemini Live API has handled model transitions differently: endpoint path changes accompanied some version shifts, requiring client reconnect logic rather than just a config update. Providers routing through DashScope or similar inference gateways have seen model retirement that affected cost calculation before the public deprecation notice landed.

The January 2027 deadline also coincides with the period when the new GPT Transcribe and GPT Live Transcribe models (launched July 28) are expected to see significant adoption. Operators who were previously using gpt-4o-mini-transcribe for synchronous file transcription now have two migration options: the new snapshot (gpt-4o-mini-transcribe-2025-12-15) for backward compatibility, or gpt-transcribe/gpt-live-transcribe if they want to move to the more capable transcription-specific stack that landed this week.

What TheRouter users should audit now

Any routing configuration that references the deprecated model IDs by name should be updated before January 2027. The practical deadline is earlier — the holiday period in December is a poor time to discover production audio traffic failing.

For teams running TheRouter with OpenAI audio providers configured, the model-name field in your provider settings is the primary place to audit. Test the replacement model IDs in staging before committing the routing table change to production, particularly for latency-sensitive realtime sessions where connection establishment time matters.