OpenAI sets 2027 shutdowns for its realtime and Whisper models
The gpt-realtime and audio lines retire Jan 20; whisper-1 and the gpt-4o transcribers Feb 26. Replacements are named — and Gemini just shipped a rival.
Copy markdown
OpenAI's realtime line has a January clock
gpt-realtime, gpt-4o-realtime, gpt-audio, gpt-4o-audio, and the -mini variants all switch off Jan 20, 2027. Voice agents built on those endpoints stop working unless you move to gpt-realtime-2.1, gpt-realtime-2.1-mini, or gpt-audio-1.5.
Whisper and the gpt-4o transcribers follow in February
whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, and gpt-4o-transcribe-diarize retire Feb 26, 2027. OpenAI routes you to gpt-transcribe for batch and gpt-live-transcribe for streaming; a pinned whisper-1 is now a dated liability.
The Assistants API is already gone
Its one-year sunset hit Aug 26, 2026 — the endpoint is removed, not just deprecated. If threads-and-runs code is still in production it is failing now; the Responses API and Conversations API are the drop-in replacements.
Gemini just shipped the rival transcriber
gemini-3.5-transcribe (batch, 85+ languages) and gemini-3.5-transcribe-live (streaming over WebSockets) went GA Aug 26, giving you a second vendor for the exact workload OpenAI is churning under you.
What to do before January
Grep your codebase for the retiring model IDs now, put every audio and transcription call behind one adapter, and smoke-test gpt-realtime-2.1 and gpt-transcribe (or Gemini's) while there is slack — not the week the endpoints start returning 400s.
Elsewhere: OpenAI's mTLS identity federation hit GA
Since Aug 29 you can register X.509 workload certificates and identity providers directly in the console with role-based access, replacing long-lived API keys for server-to-server calls.