# Cartesia vs OpenAI TTS (2026): A monthly allowance or a bill per character?

By Samuel Oram-Jones

Cartesia's $49 Startup plan buys 1.25 million credits a month, and because Sonic text to speech costs one credit per character, that is 1.25 million characters of speech, about 1,667 minutes by Cartesia's count. The same 1.25 million characters on OpenAI's `tts-1` costs $18.75 with no plan around it. As of November 2026, the cheaper line is also the one on a clock, since OpenAI says `tts-1`, `tts-1-hd`, and `gpt-4o-mini-tts` shut down on 6 January 2027.

Bland Speech takes a third route: $0.04 per 1,000 characters ($40 per million), pay as you go with no monthly plan, and Bland Speech v3 ranked first of 13 AI voices in Design Arena's Audio Realism Bench results of 4 August 2026.

## Which is better, Cartesia or OpenAI TTS?

Cartesia is the better pick for a new voice agent, because OpenAI's speech endpoint models all shut down on 6 January 2027 and Cartesia's Sonic 3.6 is a generally available model with 8 kHz mu-law output and a first-audio claim of under 90ms. OpenAI's endpoint is the better pick only for a prototype that has to run until the shutdown on a team already holding an OpenAI key.

For a phone agent that has to sound like a person, the pick is Bland Speech, though it does not lead on price or published latency.

| | Bland Speech | Cartesia | OpenAI TTS |
|---|---|---|---|
| Current model | Bland Speech v3 | Sonic 3.6, 44 languages | `tts-1`, `tts-1-hd`, and `gpt-4o-mini-tts`, shutting down 6 January 2027 |
| Billing | $40 per million characters, pay as you go | 1 credit per character on plans: Free $0, Pro $5, Startup $49, Scale $299 a month | `tts-1` $15 and `tts-1-hd` $30 per million; `gpt-4o-mini-tts` by token |
| Published latency | 199ms to first audio | Under 90ms to first audio, as Cartesia words it | A relative claim: `tts-1` is lower latency than `tts-1-hd` |
| Phone audio | mu-law at 8 kHz, PCM 8 to 48 kHz | `pcm_mulaw` and `pcm_alaw`, PCM and others, from 8 kHz up | mp3, opus, aac, flac, wav, 24 kHz pcm |
| Streaming | HTTP chunks and WebSocket | Bytes, SSE, and WebSocket; only WebSocket takes partial text | HTTP chunked; WebSocket only on the Realtime API |
| Cloning | Instant from about 10 seconds; professional from 30 minutes or more | Instant from 10 seconds; professional on Startup and above | Eligible customers only, samples of 60 seconds or less |
| Realism on 4 August 2026 | 1st of 13 at 1365 | Sonic 3.5 6th at 1096; Sonic 3.6 released after | GPT-4o Mini TTS 9th at 1011; GPT Realtime 2 4th at 1119 |

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

- **Billing:** Cartesia sells a monthly allowance of credits and counts each text character as one credit. OpenAI and Bland bill for what is used.
- **Published latency:** Time to first audio is how long a caller waits before the first sound arrives. Each figure is the company's own statement, measured by its own method, so the three are not a race.
- **Realism on 4 August 2026:** Position and Elo score in Design Arena's Audio Realism Bench results of that date. Elo is a rating built from head-to-head votes, here from listeners picking the clip that sounds more like a person.

</details>

## Where do Cartesia and OpenAI TTS fall short for a voice agent?

Cartesia falls short on bill predictability, since each plan caps credits, and OpenAI TTS falls short on lifespan, since its three speech models end on 6 January 2027.

### Cartesia

The plan ladder runs Free at 20,000 credits, Pro at $5 for 100,000, Startup at $49 for 1.25 million, and Scale at $299 for 8 million, with Enterprise custom. Credits roll over up to twice the monthly allotment. Overages are opt-in and billed when the plan renews, at $65 per million credits on Pro, $45 on Startup, and $38 on Scale, and with overages off Cartesia blocks new requests once credits run out. Free is for evaluation and personal use, and the commercial license starts on Pro.

Extras draw down the same pool. Each break tag counts as one credit, and adding an accent to a voice costs 225 credits. Emotion controls are in beta and take English only, and speed and volume do not apply to Professional Voice Clones. Cartesia has retired model ids on a schedule too, with the original `sonic` models ending on 1 June 2026 and `sonic-2` and `sonic-turbo` carrying a sunset date of 20 October 2026.

<div class="cmp-proscons" data-rival="cartesia"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>First audio in under 90ms, as Cartesia words it</li><li>44 languages on Sonic 3.6</li><li>Mu-law and A-law at 8 kHz, plus word-level timestamps on WebSocket and SSE</li><li>SOC 2 Type II, and on-prem, air-gapped, or own-VPC deployment under enterprise contracts, as Cartesia words it</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>Monthly credit plans with overages billed at renewal</li><li>Break tags and added accents draw down the same credit pool</li><li>Professional cloning needs the $49 Startup plan</li><li>Emotion controls in beta and English only</li></ul></div></div>

A poor fit for:

- A team that wants one rate with no monthly allowance to manage
- A script heavy with timed pauses, since each break tag is billed as a credit

### OpenAI TTS

OpenAI's speech endpoint is a single call that returns audio in mp3, opus, aac, flac, wav, or 24 kHz pcm, and it bills per use. The cost of that simplicity is the date: OpenAI announced on 1 October 2026 that its three speech models were deprecated, with the replacement, `gpt-realtime-2.1-mini`, on the Realtime API, which OpenAI says is "a change of API workflow".

The endpoint has no mu-law format, so a phone integration receives 24 kHz pcm or a compressed format, and OpenAI's Realtime API is where `audio/pcmu` output appears. Server-sent events are not supported on `tts-1` or `tts-1-hd`. Custom voices from recordings are limited to eligible customers, and OpenAI's usage policies require a clear disclosure to end users that the voice is AI-generated.

<div class="cmp-proscons" data-rival="openai"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>`tts-1` at $15 per million characters, billed per use</li><li>A plain-language `instructions` field on `gpt-4o-mini-tts`</li><li>SOC 2 Type 2 for its API and a Business Associate Agreement offered</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>All three speech models shut down on 6 January 2027</li><li>No mu-law on the speech endpoint</li><li>The replacement is on a different API</li><li>Custom voices only for eligible customers</li></ul></div></div>

A poor fit for:

- Any build that has to run past 6 January 2027 without a rewrite
- A phone line that wants mu-law from the speech endpoint

Other ways out of OpenAI's endpoint are in [OpenAI TTS alternatives](/blog/openai-tts-alternatives), and Cartesia's own list is in [Cartesia alternatives](/blog/cartesia-alternatives).

## What do Cartesia, OpenAI TTS, and Bland Speech cost?

For 1.25 million characters a month, Cartesia's Startup plan costs $49, OpenAI's `tts-1` costs $18.75, `tts-1-hd` costs $37.50, and Bland Speech costs $50.

At 10 million characters a month, Cartesia's Scale plan covers 8 million with its $299 allotment and bills the other 2 million at $38 per million, which is $375 in total with overages on. OpenAI's `tts-1` costs $150 and `tts-1-hd` $300, and Bland Speech costs $400. Used in full, Cartesia's Startup plan works out to $39.20 per million credits and Scale to $37.38, so the plans land close to Bland Speech's $40 once the allowance is spent, and cost more per character when it is not.

Bland Speech costs more per character than OpenAI's `tts-1` and `tts-1-hd`, and a team that prices only per character will choose OpenAI for the months it lasts. Phone minutes, taxes, and the language model that writes the text come on top of every figure here. Cartesia charges $0.014 per minute for telephony on a Cartesia-provided number, with language model usage billed separately.

## Which sounds more human, Cartesia, OpenAI TTS, or Bland Speech?

In Design Arena's Audio Realism Bench results of 4 August 2026, Bland Speech v3 ranked first of 13 AI voices at 1365, against a human recording at 1500, OpenAI's GPT Realtime 2 ranked 4th at 1119, Cartesia's Sonic 3.5 ranked 6th at 1096, and OpenAI's GPT-4o Mini TTS ranked 9th at 1011.

Cartesia's Sonic 3.6, released on 27 August 2026, is not scored in those results, so Cartesia's current model has no 4 August number. Listeners in Design Arena's method hear two recordings of one transcript, blind and in random order, and pick the one that sounds more like a real person, as [Design Arena's article](https://notes.designarena.ai/audio-realism-benchmark-measuring-realism-in-audio-models/) describes.

Cartesia's Sonic 3.5 sat 85 points above OpenAI's GPT-4o Mini TTS, the scored model from the speech endpoint, and 23 points below GPT Realtime 2, which sits on the Realtime API.

## Why do people comparing Cartesia and OpenAI TTS choose Bland Speech?

People comparing Cartesia and OpenAI TTS choose Bland Speech because Bland Speech keeps OpenAI's request shape and pay-per-use billing, returns phone-ready mu-law as Cartesia does, and Bland Speech v3 took first place in the 4 August 2026 realism results.

Bland publishes 199ms to first audio, which is slower than Cartesia's under-90ms claim, and the two figures come from different methods. A client written for OpenAI's `/audio/speech` changes its base URL, API key, and model to route through Bland, with `speed` limited to 1.0 and `instructions` ignored, and the native endpoints, on the [text to speech API](/text-to-speech-api) page, are `POST /v2/tts` and a WebSocket at `/v2/tts/ws` that takes text as a language model writes it. Bland Speech has a first-party Pipecat service and a LiveKit Agents plugin for Python. The move from OpenAI is laid out in [how to migrate from OpenAI TTS](/blog/migrate-from-openai-tts).

An instant clone needs one clean sample of about 10 seconds, a professional clone is fine-tuned on 30 minutes or more of audio after Bland's consent and rights review, and every clone is private to the customer's organization. Bland is SOC 2 Type II audited and HIPAA compliant with a signed BAA.

<div class="cmp-proscons cmp-proscons-own"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>Ranked first of 13 AI voices in Design Arena's results of 4 August 2026</li><li>One standard rate with no monthly allowance to manage</li><li>8 kHz mu-law output and an OpenAI-compatible endpoint</li><li>Native HTTP and WebSocket endpoints for streamed text</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>More than a hobby project needs, because Bland Speech is built for production calls</li></ul></div></div>

The shutdown date is what decides this pair: a phone agent started now on OpenAI's endpoint gets rewritten within weeks, and one started on Cartesia buys a credit plan, so Bland Speech, on an OpenAI-shaped endpoint with no plan and the top realism score of 4 August 2026, is the one to build on. More matchups sit in the [text to speech hub](/compare/text-to-speech) and the [rival vs rival family](/compare/text-to-speech/rival-vs-rival), and Bland Speech itself is at [bland.ai/speech](/speech).

## Frequently asked questions

### Is Cartesia better than OpenAI TTS?

For a new voice agent, Cartesia is better, because OpenAI's `tts-1`, `tts-1-hd`, and `gpt-4o-mini-tts` shut down on 6 January 2027. Cartesia's Sonic 3.6 is generally available, outputs 8 kHz mu-law, and carries a first-audio claim of under 90ms.

### What happens to OpenAI TTS projects after 6 January 2027?

The three speech endpoint models stop working. OpenAI points developers to `gpt-realtime-2.1-mini` on its Realtime API and says the move is a change of API workflow, because that model is not supported on `/v1/audio/speech`.

### Which is cheaper, Cartesia or OpenAI TTS?

OpenAI's `tts-1` is cheaper per character at $15 per million, against Cartesia's $37.38 to $50 per million depending on plan and how much of the allowance gets used. Cartesia adds plan tiers, and OpenAI bills per use with no plan.

### Do Cartesia and OpenAI TTS output audio for phone calls?

Cartesia does directly, and OpenAI's speech endpoint does not. Cartesia offers `pcm_mulaw` and `pcm_alaw` at 8,000 Hz, which Cartesia names for Twilio and North American telephony. OpenAI's speech endpoint returns mp3, opus, aac, flac, wav, or pcm at 24 kHz. Bland Speech outputs 8 kHz mu-law.

### Can Cartesia or OpenAI TTS clone a voice?

Both can, with conditions. Cartesia clones instantly from 10 seconds of audio and trains professional clones on its Startup plan and above. OpenAI offers custom voices to eligible customers from samples of 60 seconds or less. Bland Speech clones from one clean sample of about 10 seconds.

### Do Cartesia and OpenAI support SOC 2 and a BAA?

Both do. Cartesia states it is SOC 2 Type II certified and puts BAAs and DPAs on its Enterprise plan. OpenAI has an independent SOC 2 Type 2 examination for its API and offers a Business Associate Agreement. Bland is SOC 2 Type II audited and HIPAA compliant with a signed BAA.

### Is Bland Speech cheaper than Cartesia?

Bland Speech costs $40 per million characters, which is close to Cartesia's $39.20 on Startup and $37.38 on Scale when the allowance is used in full, and less than Cartesia's $50 on Pro. Cartesia bills a monthly plan, and Bland Speech has none.
