Since November 2023, speech from OpenAI has been one HTTP request: text goes in, an MP3 comes out. That ends on January 6, 2027. On October 1, 2026, OpenAI put tts-1, tts-1-hd and gpt-4o-mini-tts on its deprecation list.
The named replacement, gpt-realtime-2.1-mini, runs only on the live Realtime API and "isn't a supported model for /v1/audio/speech", says OpenAI's own text to speech guide. Creative AI News checked nine open source apps with OpenAI speech, found seven on a retiring model by default and summed it up: "Changing one model string fixes nothing."
What changes on the Realtime API?
- Audio arrives as 24 kHz PCM chunks inside events. You build the file yourself.
- Your text becomes instructions to a reasoning model (OpenAI's example starts with "Say exactly the following:"), so check names and prices by ear.
- Fable, Onyx and Nova, three of the six launch voices, are gone. A podcast narrated by Onyx gets a new voice.
Where should you move?
- Your app talks live, as a voice agent or on a phone line: use the Realtime API.
- You want to keep your /v1/audio/speech code: Kokoro-FastAPI serves the same request on your own machine. It speaks 8 languages and cannot clone a voice (more in our comparison of free local text to speech).
- You make audio files, such as narration, voiceovers, audiobooks, lessons or game dialogue: Demodokos Foundry makes them on your own PC.
What does an hour of speech cost?
List prices on November 21, 2026, at 920 characters a minute (the speaking rate of our test take below):
| Option | List price | One hour |
|---|---|---|
OpenAI tts-1, until January 6 | $15 per 1M characters | $0.83 |
OpenAI gpt-realtime-2.1-mini | $20 per 1M audio output tokens | $1.44, plus the text you send |
| ElevenLabs Eleven v4 (API) | $0.08 per 1,000 characters | $4.42 |
| Amazon Polly Neural | $16 per 1M characters | $0.88 |
| Demodokos Foundry Professional | $39.20 a month billed yearly, $49 billed monthly | Flat price, unlimited generation |
For a few hours a month the cloud costs less. At $49 billed monthly, Foundry Professional costs less than the Realtime API from about 34 hours of speech a month, and less than Eleven v4 from about 11. Retakes cost nothing extra.
How does a tts-1 script move to Foundry?
Demodokos Foundry is a Windows program that makes speech, music, dubbing and transcripts on your own graphics card. On the Professional plan it has its own local HTTP API and a command line:
import requests
r = requests.post("http://localhost:19847/speech/direct", json={
"voice_name": "Harbor Narrator", # a voice you designed or cloned in Foundry
"text": text,
"emotion": "narrator",
"language": "en",
"output_format": "mp3",
}, timeout=1800)
print(r.json()["segments"][0]["outputPath"])
We sent it the first 999 words of Dracula, 5,371 characters (OpenAI's endpoint takes 4,096 per request). The MP3 with 5 minutes 49 seconds of speech was on disk 60 seconds later. The take, unedited:
What do you get on your own PC?
- No provider can retire your voice. The voices you design or clone stay in your project.
- Unlimited generation and no price per character. Professional has 38 emotions and speaking styles.
- Local inference. Local storage. Encrypted connection. No creative content uploaded to Demodokos for cloud AI processing, and voice cloning runs locally.
- When whisper-1 retires on February 26, 2027, Foundry's Dictation transcribes on the same PC.
Frequently asked questions
When does OpenAI shut down tts-1?
On January 6, 2027, together with tts-1-hd, gpt-4o-mini-tts and its snapshots gpt-4o-mini-tts-2025-03-20 and gpt-4o-mini-tts-2025-12-15.
Can I still use /v1/audio/speech after January 6, 2027?
As of November 21, 2026, OpenAI has named no model for it after that date, and gpt-realtime-2.1-mini does not run on it.
Will my OpenAI voice still be available?
Only if it is one of the ten Realtime voices: alloy, ash, ballad, coral, echo, sage, shimmer, verse, marin and cedar.
Is OpenAI also shutting down Whisper?
Yes. whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize shut down on February 26, 2027. OpenAI names gpt-transcribe and gpt-live-transcribe as replacements.
Try the local API for 7 days
Point one tts-1 script at your own PC before January 6. The API comes with Professional.
Start the free trial