OpenAI text to speech for AI agents: pricing, shutdown dates and setup

What is OpenAI text to speech?

OpenAI's Audio API has a speech endpoint that reads text aloud in one of 13 built-in voices, such as Alloy, Coral, Marin and Cedar, and a transcription endpoint that turns a recording into text. Both are reached with an ordinary OpenAI API key.

The speech endpoint takes up to 4,096 characters per request and returns MP3, Opus, AAC, FLAC, WAV or raw PCM. With gpt-4o-mini-tts you can also pass instructions for tone, accent and pace. OpenAI says the voices are optimized for English, though they follow Whisper's language coverage.

gpt-4o-mini-tts and Whisper are being shut down

OpenAI's deprecations page says tts-1, tts-1-hd and both snapshots of gpt-4o-mini-tts shut down on January 6, 2027. The replacement, gpt-realtime-2.1-mini, isn't supported on the speech endpoint: it works through the Realtime API over WebRTC, WebSocket or SIP. Developers on OpenAI's forum point out that a realtime model has to be prompted to read text exactly, which a speech endpoint never needed.

Transcription follows on February 26, 2027: whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize all shut down, replaced by gpt-transcribe and gpt-live-transcribe.

Rakazo speaks with gpt-4o-mini-tts and transcribes with gpt-4o-mini-transcribe today. Until Rakazo moves to the newer models, OpenAI voice in Rakazo stops working on those dates.

From OpenAI's deprecations page, checked November 2026.

OpenAI TTS pricing

OpenAI has no free tier for audio; you prepay API credit. The older speech models are billed per character, and the newer ones per token, which is why people find the pricing confusing. gpt-4o-mini-tts charges for the text you send and, much more, for the audio it returns. OpenAI used to estimate that at about $0.015 a minute; its pricing page no longer gives a per-minute figure, and some developers report bills above that estimate.

From OpenAI's API pricing page, checked November 2026. Hourly speech figures assume about 54,000 characters an hour; the gpt-4o-mini-tts figure uses OpenAI's earlier estimate of $0.015 a minute.

What a talking Rakazo bot costs on OpenAI

Rakazo calls OpenAI with your key, so OpenAI bills you and Rakazo adds nothing. The voice key is stored separately from any OpenAI key you use for models, but both draw on the same OpenAI account if you use the same project.

The worked example is one bot that reads about 30 minutes of replies aloud each day and transcribes about 30 minutes of what you say, every day for a month: 15 hours of speech (about 810,000 characters) and 15 hours of transcription. On OpenAI that is about $13.50 of speech and $2.70 of transcription: about $16 a month, the cheapest hosted option in Rakazo, until the shutdown.

Usage prices from each provider's pricing page, checked November 2026. Where speech is bought as a monthly plan, as with Cartesia, the cost shown is the fee for the smallest plan that covers it. Rakazo doesn't charge for voice: you connect your own key and the provider bills you.

OpenAI TTS vs ElevenLabs and other voice APIs

OpenAI is cheap and simple, with a small fixed set of voices. ElevenLabs and Fish Audio sound more natural to most listeners and let you use cloned voices; Cartesia is faster to start talking. OpenAI's custom voices exist but are limited to eligible customers through sales, and Rakazo doesn't use them.

Prices in US dollars from each provider's pricing page, checked November 2026. An hour of speech is taken as about 54,000 characters (150 words a minute); real voices run a little faster or slower. Fish Audio bills per UTF-8 byte, which is one byte per character in English.

What you can do with OpenAI in Rakazo

OpenAI both speaks and transcribes, so it covers every voice feature in Rakazo. The voice and calls guide shows where each control lives.

• Talk instead of typing. Hold the mic in the composer, speak, and let go to send. When Rakazo transcribes on the server, as it does in a phone browser or the desktop app, it uses gpt-4o-mini-transcribe.
• Hear replies. Choose Speak this reply on any message, or turn on Read replies aloud in a bot's settings. Replies are spoken with gpt-4o-mini-tts.
• Call a bot. Choose Call at the top of a chat to talk in real time. You can cut in while the bot is speaking, and the transcript stays in the chat.
• Get called. Ask "call me when the drafts are ready" and the bot rings you. A bot may also call if you've missed the same reminder several times.
• Give each bot its own voice. Pick from OpenAI's 13 voices per bot, so different bots are easy to tell apart.
• Send voice notes from chat apps. A voice note sent to a bot through Telegram, for example, is transcribed before the bot reads it. See messaging apps.

When to choose OpenAI

Choose OpenAI when you already have an OpenAI account and want voice working in a few minutes, with clear voices and solid transcription from one key at a low price. Because its speech and transcription models are scheduled to shut down, treat it as a convenient start rather than a long-term pick, until Rakazo moves to OpenAI's newer models.

ElevenLabs also speaks and transcribes, with more natural voices and support for your own cloned ones. Cartesia is built for fast, interruptible calls but only speaks. Fish Audio offers cloning and expressive speech at a similar price to OpenAI. If you like OpenAI's API shape but want audio on your own hardware, a self-hosted server such as Kokoro or Whisper speaks the same protocol. Server (Mac) needs no key when your server runs on a Mac. The comparison puts them side by side.

Limits and gotchas

• Rakazo speaks with gpt-4o-mini-tts, which OpenAI shuts down on January 6, 2027, and transcribes with gpt-4o-mini-transcribe, which follows on February 26, 2027. Until Rakazo moves to newer models, OpenAI voice stops working on those dates.
• The models are fixed. You can't choose tts-1, a newer model or a different transcription model from Rakazo.
• Rakazo offers OpenAI's 13 built-in voices. It doesn't send tone instructions, and custom voices made in your OpenAI account aren't supported.
• Each sentence is generated as a finished clip and played in order, rather than streamed, so there is a short pause before the bot starts talking.
• OpenAI keeps speech requests for up to 30 days for abuse monitoring and doesn't train on API data by default. Transcription requests aren't retained. Zero data retention needs OpenAI's approval.
• OpenAI's usage policies ask you to tell listeners the voice is AI-generated. Keep that in mind if a bot's voice reaches people outside your team.
• Calls and hold-to-talk are a server setting. On a self-hosted server they stay off until the admin turns them on.

How to set it up

Anyone can do this in their own settings, on Rakazo Cloud or a self-hosted server. The key belongs to your account, and it is stored separately from any OpenAI key you use for models, so paste it here even if you already added it elsewhere.

1. Create an API key in the OpenAI platform for a project with audio access.
2. In Rakazo, open Settings → Voice, choose OpenAI as the Provider, paste the API key and choose Connect.
3. Pick a Voice, such as Marin or Cedar, and use Hear a sample to check it.
4. Optionally, open a bot's settings, expand Advanced, turn on Read replies aloud and give it its own Voice.
5. If you run your own server and want calls and hold-to-talk, set PHONE_CALLS_ENABLED=true on the API and worker.