# ElevenLabs vs OpenAI TTS (2026): What a shutdown date does to the choice

By Samuel Oram-Jones

OpenAI's text to speech endpoint has an end date. The models behind it, tts-1, tts-1-hd, and gpt-4o-mini-tts, are deprecated and "will shut down on January 6, 2027", OpenAI says, a date OpenAI announced on 1 October 2026. A developer starting a voice agent on it would be building on a model with an end date already set, and a developer already running on it has a migration to schedule.

OpenAI names a replacement, gpt-realtime-2.1-mini, and says plainly that it "isn't a supported model for `/v1/audio/speech`", which makes the move a change of API workflow and not a model swap. ElevenLabs' newest models, Eleven v4 and Eleven v4 Turbo, launched on 28 September 2026. A team that needs a voice on a phone line should weigh Bland Speech beside both.

## Which is better, ElevenLabs or OpenAI TTS?

As of November 2026, ElevenLabs is the better choice for a new text to speech build, because every OpenAI model on the speech endpoint has a shutdown date and ElevenLabs' newest models launched in September. OpenAI TTS still has the lower per-character price, $15 per million characters on tts-1 against $40 to $80 on ElevenLabs' current API rates, but a low price on a model that stops on January 6, 2027 is a short-lived price.

| Feature | Bland Speech | ElevenLabs | OpenAI TTS |
|---|---|---|---|
| Status | Bland Speech v3 is current | Eleven v4 and Eleven v4 Turbo, launched 28 September 2026 | tts-1, tts-1-hd, and gpt-4o-mini-tts shut down January 6, 2027, with gpt-realtime-2.1-mini named as the replacement |
| Billing | $0.04 per 1,000 characters, pay as you go | Dollars per 1,000 characters on the API, with plans from $6 a month | $15 per million characters on tts-1, $30 on tts-1-hd, and tokens on gpt-4o-mini-tts |
| First audio as published | 199ms | About 100ms median on v4 Turbo, excluding application and network latency | tts-1 "provides lower latency" than tts-1-hd |
| Streaming | HTTP chunked and WebSocket | HTTP chunked and WebSocket, with v4 on a separate dialogue endpoint | HTTP chunked on the speech endpoint, WebSocket and WebRTC on the Realtime API |
| Custom voice from a recording | One clean sample of about 10 seconds | 10 seconds on v4, from the $6 Starter plan | Limited to eligible customers through sales, with a consent recording |
| Phone audio | Mu-law at 8 kHz, PCM from 8 to 48 kHz | Mu-law, A-law, and PCM at 8 kHz among other formats | MP3, Opus, AAC, FLAC, WAV, and 24 kHz PCM on the speech endpoint |
| Realism result, 4 August 2026 | Ranked 1st of 13 at 1365 | Eleven v3 ranked 12th at 938 | GPT Realtime 2 ranked 4th at 1119, GPT-4o Mini TTS 9th at 1011 |

<details><summary>What these rows mean</summary>

- **Status:** Whether the model a team would build on today is still sold. A shutdown date means requests to that model stop working on that day. The Realtime API that carries OpenAI's named replacement is a connection that stays open for a whole conversation, where the speech endpoint answers one request at a time.
- **Billing:** A token is a chunk of text or audio that OpenAI counts instead of characters. The tts-1 and tts-1-hd models bill characters, and gpt-4o-mini-tts bills $0.60 per million text tokens in and $12 per million audio tokens out.
- **Phone audio:** Mu-law is a compressed format that phone networks use at 8 kHz. OpenAI's speech endpoint offers other formats and no mu-law, and OpenAI's Realtime API accepts an `audio/pcmu` output setting, which is the same G.711 mu-law format.
- **Realism result:** The Elo score from Design Arena's Audio Realism Bench results of 4 August 2026, where listeners pick which of two clips sounds more like a real person.

</details>

## What does ElevenLabs offer a team leaving OpenAI TTS?

ElevenLabs offers a team leaving OpenAI TTS three things the speech endpoint doesn't: audio tags that direct a performance, a cloning ladder that starts on a $6 plan, and output formats made for phone lines. Audio tags such as [whispering], [shouting], and [laughing] work on Eleven v4, where OpenAI's gpt-4o-mini-tts takes a plain-language `instructions` prompt such as "Speak in a cheerful and positive tone." ElevenLabs returns mu-law and A-law at 8 kHz and names mu-law as the format Twilio audio inputs use.

The catch is where the newest models run. ElevenLabs says its standard text to speech WebSocket "does **not** support the `eleven_v3` or `eleven_v4` models", so Eleven v4 goes through the Text to Dialogue API and Eleven v4 Turbo through the Text to Dialogue WebSocket. A team porting an OpenAI integration still has a design decision to make.

### ElevenLabs pricing

ElevenLabs charges $0.08 per 1,000 characters on Eleven v4 and $0.04 on Eleven v4 Turbo and the Flash models. Plans are Starter at $6 a month, Creator at $22, Pro at $99, Scale at $299, and Business at $990, with Enterprise custom, and the Starter plan includes 75,000 characters at the v4 rate. A free account is limited to non-commercial use, so a production agent starts at Starter. One million characters costs $40 on Eleven v4 Turbo and $80 on Eleven v4 at the API rates, against $15 on OpenAI's tts-1.

ElevenLabs offers Business Associate Agreements to qualifying healthcare enterprises and holds SOC 2 Type 2. A professional voice clone needs about 30 minutes of audio and a Creator plan or above.

<div class="cmp-proscons" data-rival="elevenlabs"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>Newest models launched 28 September 2026</li><li>Audio tags for whispering, shouting, and laughing</li><li>Instant cloning from 10 seconds on a $6 plan</li><li>Mu-law and A-law output at 8 kHz</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>$40 to $80 per million characters against tts-1's $15</li><li>Eleven v3 and v4 don't run on the standard text to speech WebSocket</li><li>Free accounts are non-commercial</li></ul></div></div>

ElevenLabs is a poor fit for a team that chooses on list price per character alone, since OpenAI's tts-1 costs $0.015 per 1,000 characters.

## What did OpenAI TTS do well, and what does the shutdown change?

OpenAI TTS did well on simplicity and price, and the shutdown changes what a new project can rely on. The speech endpoint has no plans to choose between: tts-1 costs $15 per million characters, tts-1-hd costs $30, and gpt-4o-mini-tts steers delivery through a natural-language prompt. OpenAI says the API has had an independent SOC 2 Type 2 examination and offers a Business Associate Agreement, and its usage policies require a clear disclosure to end users that the voice is AI-generated.

OpenAI's speech endpoint has limits a voice agent feels. OpenAI says tts-1 "provides lower latency, but at a lower quality than the `tts-1-hd` model." The `sse` stream format isn't supported on tts-1 or tts-1-hd. The speech endpoint's formats are MP3, Opus, AAC, FLAC, WAV, and PCM at 24 kHz, with no mu-law, so audio bound for a phone line needs a conversion step. OpenAI says voices are "currently optimized for English" and that language support generally follows Whisper's.

### OpenAI TTS cloning and the replacement

Custom voices from recordings are "limited to eligible customers", up to 20 per organization, from samples of 60 seconds or less, with a consent recording of the voice actor reading one of OpenAI's approved phrases. A team without an eligible account can create up to 100 voices from a text prompt instead, and no recording is involved.

The replacement is a different product. gpt-realtime-2.1-mini runs over the Realtime API on WebRTC or WebSocket, and bills audio at $10 per million tokens in and $20 per million out. The Realtime API accepts `audio/pcmu` output in its session settings. A migration moves from a request-and-response endpoint to a session that stays open, and it needs new code.

<div class="cmp-proscons" data-rival="openai"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>No plans, pay per use</li><li>tts-1 at $15 per million characters</li><li>A plain-language instructions prompt on gpt-4o-mini-tts</li><li>SOC 2 Type 2, with a BAA offered</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>tts-1, tts-1-hd, and gpt-4o-mini-tts shut down January 6, 2027</li><li>The replacement isn't supported on the speech endpoint</li><li>No mu-law on the speech endpoint</li><li>Recording-based custom voices limited to eligible customers</li></ul></div></div>

OpenAI TTS is a poor fit for a project that has to run past the first week of January 2027 on its current endpoint.

## Which sounds more human, ElevenLabs or OpenAI TTS?

On Design Arena's Audio Realism Bench results of 4 August 2026, OpenAI's GPT Realtime 2 ranked 4th of 13 AI voices at 1119 and GPT-4o Mini TTS ranked 9th at 1011, while ElevenLabs' Eleven v3 ranked 12th at 938. Elo is a rating that goes up each time listeners pick a voice as more human, so OpenAI's two models scored 181 and 73 points above Eleven v3.

Two cautions apply. Eleven v4 and Eleven v4 Turbo launched on 28 September 2026, after those results, so the 4 August scores don't cover ElevenLabs' current models. GPT Realtime 2 is a different model from gpt-realtime-2.1-mini, the replacement OpenAI names for the speech endpoint, and gpt-realtime-2.1-mini wasn't in the results at all. In those results Bland Speech v3 ranked 1st of 13 at 1365, 427 points above Eleven v3, with the human recording at 1500. Design Arena set out the method and the scores in its [Audio Realism Bench article](https://notes.designarena.ai/audio-realism-benchmark-measuring-realism-in-audio-models/).

## What does Bland Speech offer a team leaving OpenAI TTS?

Bland Speech offers a team leaving OpenAI TTS an endpoint that implements OpenAI's `/audio/speech` contract, and Bland Speech v3, which ranked first of 13 AI voices in Design Arena's Audio Realism Bench results of 4 August 2026, at 1365. A client already written for OpenAI's speech API changes its base URL, API key, and model to route speech through Bland.

That compatibility has stated limits. The `speed` field accepts only 1.0, `instructions` is accepted and ignored, OpenAI model IDs such as `tts-1` and `gpt-4o-mini-tts` aren't aliased, and input is capped at 4,096 characters. Formats are MP3, Opus, AAC, FLAC, WAV, and PCM at 24 kHz. Teams that want mu-law at 8 kHz, performance tags such as [laughs] and [say warmly], and a WebSocket that takes text as a language model writes it use Bland's native endpoints, covered on the [text to speech API](/text-to-speech-api) page.

Bland Speech costs $0.04 per 1,000 characters, pay as you go, with 50,000 free characters once for a new account, which is $40 per million against tts-1's $15. Bland publishes 199ms to first audio. An instant clone needs one clean sample of about 10 seconds, and every clone is private to the customer's organization, with the customer confirming the voice is theirs to use. A professional clone is fine-tuned on 30 minutes or more of audio and goes through Bland's consent and rights review.

Bland is SOC 2 Type II audited and HIPAA compliant with a signed BAA. Enterprise pricing is custom, set with Bland's sales team.

<div class="cmp-proscons cmp-proscons-own"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>An endpoint that follows OpenAI's `/audio/speech` contract</li><li>Bland Speech v3 ranked 1st of 13 in the 4 August 2026 realism results</li><li>One standard rate with no monthly plan</li><li>Mu-law at 8 kHz on the native endpoint</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>Built for production calls, which is more than a weekend prototype needs</li></ul></div></div>

The [OpenAI TTS migration guide](/blog/migrate-from-openai-tts) maps the request fields, [best OpenAI TTS alternatives](/blog/openai-tts-alternatives) lists the other options, and the [Bland Speech vs OpenAI TTS](/compare/bland-speech-vs-openai-tts) page sets the two side by side. The [text to speech hub](/compare/text-to-speech) and the [rival vs rival family](/compare/text-to-speech/rival-vs-rival) hold the rest of the series, and [Bland Speech](/speech) has the studio.

## Frequently asked questions

### Is OpenAI TTS being shut down?

Yes. OpenAI says tts-1, tts-1-hd, and gpt-4o-mini-tts "will shut down on January 6, 2027", and OpenAI announced the shutdown on 1 October 2026. OpenAI names gpt-realtime-2.1-mini on the Realtime API as the replacement.

### What replaces OpenAI TTS?

OpenAI points to gpt-realtime-2.1-mini, which runs on the Realtime API over WebRTC or WebSocket and isn't a supported model for `/v1/audio/speech`. A team can also move to a text to speech provider such as ElevenLabs or Bland Speech, and Bland's OpenAI TTS migration guide covers the Bland Speech route.

### Is ElevenLabs better than OpenAI TTS?

For a new build, ElevenLabs is the safer choice, because OpenAI's speech models end on January 6, 2027 and ElevenLabs' newest models launched on 28 September 2026. OpenAI's tts-1 is cheaper per character, at $15 per million characters against ElevenLabs' $40 to $80.

### Does OpenAI TTS output mu-law for phone calls?

Not on the speech endpoint, which returns MP3, Opus, AAC, FLAC, WAV, and PCM at 24 kHz. OpenAI's Realtime API accepts an `audio/pcmu` output setting. ElevenLabs offers mu-law and A-law at 8 kHz, and Bland Speech offers mu-law at 8 kHz.

### Can OpenAI TTS clone a voice?

Only for eligible customers through sales. OpenAI limits custom voices from recordings to eligible customers, at most 20 per organization, from samples of 60 seconds or less with a consent recording. ElevenLabs sells instant cloning from its $6 Starter plan, and Bland Speech's instant clone takes one clean sample of about 10 seconds.

### Does Bland Speech work with OpenAI's speech API clients?

Yes, with limits. Bland Speech's compatible endpoint takes a changed base URL, API key, and model, accepts `speed` only at 1.0, ignores `instructions`, and doesn't alias OpenAI model IDs.
