What changed

OpenAI notified API developers that legacy GPT Audio, Realtime and transcription model families and snapshots will be removed on 20 January 2027. Its notice names GPT-Audio 1.5, GPT-Realtime 2.1 and their mini variants as successors for most affected routes.

Why it matters

Voice and transcription applications can fail at the model-ID boundary even when their transport and application code remain unchanged. The notice makes endpoint inventory, output compatibility, latency, cost and fallback behavior immediate migration work.

What to do

Export every production audio, realtime and transcription model ID, map each to OpenAI's stated replacement, and replay consented multilingual and noisy samples. Verify WebRTC or WebSocket events, interruption handling, timestamps, tool calls, latency, cost and rollback before changing aliases.

Caveat

The shutdown date and replacement mapping are OpenAI documentation, not independent evidence that successors preserve accuracy, latency, voice behavior or response formats. Some snapshots have different replacements, and the current table must be rechecked before cutover.