OpenAI will remove tts-1 and gpt-4o-mini-tts on 6 January 2027

OpenAI’s deprecations page has a section dated 2026-10-01, “Text-to-speech models.” It says those models “will be removed from the API on January 6, 2027,” with at least three months’ notice. The named replacement for all four rows is gpt-realtime-2.1-mini.

Shutdown date	Model ID	Recommended replacement
6 Jan 2027	tts-1	gpt-realtime-2.1-mini
6 Jan 2027	tts-1-hd	gpt-realtime-2.1-mini
6 Jan 2027	gpt-4o-mini-tts-2025-03-20	gpt-realtime-2.1-mini
6 Jan 2027	gpt-4o-mini-tts-2025-12-15	gpt-realtime-2.1-mini

The text-to-speech guide still shows a one-call audio.speech.create against gpt-4o-mini-tts, with a voice name and input text, writing an MP3. The create-speech reference is that endpoint, v1/audio/speech. The tts-1 model page says to use the model with the Speech endpoint. Until 6 January 2027, that is the documented path. After that date, the deprecations table says these IDs are gone.

What the named replacement actually supports

The gpt-realtime-2.1-mini page describes a realtime voice model. It “supports audio and text inputs over WebRTC, WebSocket, or SIP connections.” Its endpoint table marks v1/realtime as supported. Speech generation (v1/audio/speech), Chat Completions, Responses, and Batch are all “Not supported,” along with transcription and translation.

Text-token prices on that page are $0.60 input, $0.06 cached input, and $2.40 output per 1 million tokens. The comparison table lists the same three text rates for GPT-Realtime Mini. Audio tokens on the mini page are a separate schedule: $10 input, $0.30 cached input, and $20 output per 1 million audio tokens. Those are OpenAI’s listed prices, not a measured bill.

Moving to this replacement means opening a realtime session instead of posting text and saving an MP3. WebRTC carries live media between peers. A WebSocket carries ordered messages on a connection you keep open. SIP is a telephony signaling path. How those transports differ for a voice agent is the practical split: a browser or phone call is usually WebRTC, and an app that already sends audio in chunks often already speaks WebSocket. Either way, the client has to hold a session, stream audio out, and handle barge-in and disconnects. A single REST call does not do that.

Another speech path that is not this endpoint

The deprecations page does not say gpt-realtime-2.1-mini is the only way to get speech from the API. gpt-audio-1.5 “can be used in the Chat Completions REST API” and lists audio as an output modality. Its endpoint table marks v1/chat/completions supported and v1/audio/speech not supported. Text tokens there are $2.50 input and $10 output per 1 million. Audio tokens are $32 input and $64 output per 1 million. The page’s own blurb calls it “the best voice model for audio in, audio out with Chat Completions.”

gpt-audio and gpt-audio-mini also mark Chat Completions as supported and v1/audio/speech as not supported. A separate deprecations section, dated 2026-07-20, schedules the gpt-audio and gpt-audio-mini families for removal on 20 January 2027, with gpt-audio-1.5 as the recommended replacement. So after 6 January 2027 the simple speech endpoint’s current model IDs are scheduled to disappear, and a Chat Completions audio model remains on the docs. That is not a drop-in change of the model field on v1/audio/speech.

Transcription models on the same page

The 26 August 2026 section lists four transcription IDs for removal on 26 February 2027: whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, and gpt-4o-transcribe-diarize. The recommended replacements in that table are gpt-live-transcribe or gpt-transcribe. That is a speech-to-text retirement, separate from the text-to-speech rows above.

What to change in code

If you call v1/audio/speech with tts-1, tts-1-hd, or gpt-4o-mini-tts, the deprecations page gives you until 6 January 2027 and names gpt-realtime-2.1-mini. That model does not implement the speech endpoint, Chat Completions, or Responses. Budget a realtime session (WebRTC, WebSocket, or SIP) for that migration. If you need a request-response speech API instead, the docs still show audio output on Chat Completions through gpt-audio-1.5, which is a different route and a different price table.