OpenAI's text-to-speech models retire on 6 January 2027

Updated 2026-10-05 · 4 min read · By the Convo360 team

What is being removed, and when

On 1 October 2026 OpenAI listed its older text-to-speech models as deprecated, to be removed from the API on 6 January 2027, with at least three months' notice. The recommended replacement is gpt-realtime-2.1-mini.

This sits alongside a wider clear out. The same deprecations page also lists transcription models deprecated in August 2026, a set of legacy audio, realtime and transcription models deprecated in July, and a group of chat models scheduled for removal on 1 April 2027. Audio has been the most actively pruned part of the OpenAI range this year.

The date is the part worth writing down. Deprecation changes nothing you can hear. Removal is when the request fails, and for a voice agent that means a caller hears silence or a dropped call rather than a slightly worse voice.

Who this actually affects

Narrower than it first sounds. This is only about turning text into speech, so the question is simply whether OpenAI is your voice engine.

• Affected: AI phone calls where the agent speaks using an OpenAI voice.
• Affected: voice notes or audio replies generated with OpenAI speech.
• Not affected: WhatsApp, Instagram and web chat bots that send text.
• Not affected: agents using Sarvam, ElevenLabs or Smallest.ai for speech, even if OpenAI is the brain.

That last line is the one people get wrong. On Convo360 the model that thinks and the model that speaks are chosen separately, so an agent can use an OpenAI model for its replies and Sarvam for its Hindi voice. In that setup, this deprecation does not touch you at all. Check the voice setting, not the brain setting.

Why the replacement is not a straight swap

The recommended destination, gpt-realtime-2.1-mini, is a realtime model rather than a plain text-to-speech endpoint. It is designed for live spoken conversation, taking audio in and producing audio out, instead of simply reading a finished sentence aloud.

For a live phone agent that is often an improvement, because a realtime model can cut the gap between the caller finishing a sentence and the agent starting to speak. Latency is what makes an AI call feel human or feel like a queue.

But it is a different shape of integration, and it bills differently. If all you need is a recorded line read aloud, for example a fixed confirmation message, a realtime conversational model is more machinery than the job requires. Decide which of the two you are actually doing before picking the replacement.

For Indian languages, consider leaving OpenAI

If your calls are in Hindi, Marathi, Tamil or Telugu, a forced migration is a good moment to ask whether OpenAI should be your voice engine at all.

Sarvam is built for Indian languages and prices speech at about three rupees per thousand characters. ElevenLabs released Eleven v4 and Eleven v4 Turbo on 28 September 2026, with higher quality and more expressive speech than the previous generation. Both tend to handle Indian names, place names and code mixed Hinglish more convincingly than a general purpose English voice.

Our comparison of Sarvam, ElevenLabs and OpenAI voices for Indian callers goes through this in detail. The short version is that for an English only agent the choice is close, and for an Indian language agent it usually is not.

How to change the voice without a bad call

On a bring-your-own-keys platform the voice is a setting on the agent, so the change is quick. What takes the time, and what matters, is listening to the result.

In Convo360 you pick the voice engine, the voice, the gender and the language on the agent itself, then place a test call. Because you hold the provider keys, swapping from OpenAI speech to Sarvam or ElevenLabs means adding that provider's key under provider keys and selecting its voice. No flow, prompt or knowledge base has to change.

Listen to a complete call rather than a sample sentence. The things that go wrong are specific and audible: a name mispronounced, a phone number read as one long string, rupee amounts spoken oddly, or a pace that sounds rushed on a weak mobile connection. Those only show up in a real conversation.

Treat this as a recurring event, not a one off

Voice models are being retired faster than chat models. In 2026 OpenAI deprecated transcription models in August, legacy audio and realtime models in July, and text-to-speech in October. Sarvam retired bulbul:v2 earlier in the year and broke every call running on it.

The lesson from that Sarvam outage is the useful one. The announcement existed, it was public, and nobody affected had read it. Businesses found out when callers complained, which is the most expensive way to learn.

So watch the deprecations pages specifically. Each provider publishes retirements separately from its product news, and the retirement page is the one that predicts an outage. Convo360 reads those pages for Meta, OpenAI, Anthropic and Google several times a day and emails the owner when something is removed or repriced, precisely so a voice does not go silent without warning.

Step by step

1. Check your voice engine, not your brainOpen each voice agent and look at the voice engine setting. If it is not OpenAI, this deprecation does not affect you.
2. Decide live conversation or read aloudA realtime model suits live calls. A fixed message read aloud may be better served by a simpler text-to-speech voice.
3. Consider an Indian language providerFor Hindi and other Indian languages, compare Sarvam and ElevenLabs rather than defaulting to another OpenAI voice.
4. Add the new provider keyIf you are changing provider, add that provider's key under provider keys. Your flows and knowledge base stay as they are.
5. Place a full test callListen to a whole conversation, checking names, phone numbers and rupee amounts, not just one sample sentence.

Example conversation

Frequently asked questions

Related reading

Available across India — see WhatsApp & AI calling automation by state and city.