OpenAI Text-to-Speech Review 2026

Published September 1, 2026 · Every fact below was read from OpenAI’s own developer documentation, pricing tables and deprecation notices on September 1, 2026, with the source linked beside the claim. Where a document could not be reached today, this review says so and gives the date it was last verified.

Quick verdict

Every price and feature claim below was read from OpenAI’s official developer pages on September 1, 2026, with the source linked beside the fact. Our standardized audio benchmark is designed but not yet running, so this review contains no listening-test claims, no naturalness verdicts and no scores.

Strongest advantages:

• Instruction-steerable delivery: the instructions parameter takes plain-language direction rather than markup, on gpt-4o-mini-tts.
• API content is contractually excluded from model training, with the speech endpoint named in the per-endpoint table.
• The strongest output-ownership language in this comparison, including an express assignment (verification note below).
• Processing is confined to the United States and Europe, which simplifies some data-residency reviews.
• Deprecations are published in a dated notice rather than discovered in production.

Important disadvantages:

• Three models, two incompatible billing units, in one table.
• No free tier of any kind for speech.
• No per-minute cost published for any speech model.
• Custom-voice licensing terms are referenced but not published.
• A disclosure duty is asserted in the guide while the policy it cites no longer contains it.
• A hard 4,096-character cap per request forces client-side chunking.
• No SSML support anywhere in the speech API.
• Half the audio price table is already scheduled for removal.

On this page

Who should use it — and who should look elsewhere

Choose OpenAI if:

• You want to direct a voice in prose rather than markup. The instructions parameter accepts plain-language direction over delivery — no SSML to learn, no tag vocabulary to maintain (TTS guide, checked September 1, 2026).
• Your compliance position depends on your content not training a model. OpenAI states it plainly: “Your data is your data. As of March 1, 2023, data sent to the OpenAI API is not used to train or improve OpenAI models (unless you explicitly opt in to share data with us)”, and the speech endpoint appears by name in the per-endpoint data table.
• You are already an OpenAI customer. One key, one bill, one SDK covering speech alongside everything else you call.
• Processing location matters. Speech processing is documented as confined to the United States and Europe.

Look elsewhere if:

• You need any free usage at all. No free speech tier is documented. Google Cloud Text-to-Speech gives four million characters a month on Standard and WaveNet (pricing, checked September 1, 2026).
• You need to quote a per-minute cost. OpenAI publishes none for speech; the per-minute column on its own pricing table is populated only for transcription.
• You need self-serve voice cloning. Custom voices require a sales conversation. ElevenLabs documents cloning from $6/month (pricing, checked September 1, 2026).
• You need SSML or long single requests. There is no SSML support, and input is hard-capped at 4,096 characters per request.

Current price, free plan, commercial rights

The short version: three speech models, two billing units, no free tier, and no per-minute figure anywhere.

Read from developers.openai.com/api/docs/pricing and cross-checked against OpenAI’s own Markdown twin of the same page, which agrees exactly, both on September 1, 2026. The section header states “Prices per 1M tokens unless noted”, and the “/ 1M characters” string is a per-cell override appearing only on the two tts-1 rows.

What this means for you: you cannot compare OpenAI’s three speech models against each other on the published figures, let alone against a competitor, without first converting tokens to characters — a conversion OpenAI does not publish for audio output. The older models are the only ones you can price per character directly.

There is no free tier for speech. None is documented on the pricing page or in the speech guide. That makes OpenAI the only provider among the ten we cover with no free path to a first generated file.

What this review covers

This review covers OpenAI’s dedicated text-to-speech models — gpt-4o-mini-tts, tts-1 and tts-1-hd — their published prices, the instruction-based control surface, custom voices and their gating, the request limits and processing regions, the published deprecation notice, and the terms governing ownership and training. It does not cover the realtime speech-to-speech models except where their deprecation touches the audio price table.

It does not evaluate how any OpenAI voice sounds. We created no account, made no API call and generated no audio. Rows a review of ours normally carries once our benchmark runs — Tested, Model tested, Voice tested, Benchmark cost — are absent rather than filled with invented values.

A verification note that governs the licensing section. OpenAI’s policy pages returned an access error to every method we tried on September 1, 2026, so the contract clauses quoted in licensing could not be re-read today. They were verified on August 11, 2026 for our flagship comparison and are reproduced here with that date attached. We would rather show you the date a clause was actually read than imply a freshness we do not have. The marketing pricing page at openai.com also returned an access error and could not be cross-checked; the developer pricing page was reachable and is the source for every figure above.

Benchmark audio status

Samples are pending. Our standardized audio benchmark is designed but not running, so this page publishes no audio and no empty player (methodology). What we verified instead is documentation: model identifiers and their status across three official pages, prices against two renderings of the same source, request caps, region scoping, the deprecation notice, and the guide’s own statements about disclosure and custom voices, all read on September 1, 2026.

Assessment by dimension

No numeric scores appear below. The judgments are qualitative and drawn from what OpenAI documents.

Models and the deprecation notice

OpenAI publishes a dated deprecation notice covering its older audio families. It records that on July 20, 2026 developers using legacy audio, realtime and transcription model families and snapshots were notified of their deprecation and removal from the API (deprecations, checked September 1, 2026). A material share of the rows on the live audio price table therefore already carries an end date.

The tts-1 lineage is described three different ways. tts-1 and tts-1-hd are absent from the “Speech generation” section of the models index, present with their own live model pages, and present with live price rows on the pricing page. All three are official and current. We report the inconsistency rather than pick the reading that makes the lineup look tidiest — but if you are choosing a model to build on today, that ambiguity is itself an argument for the current one.

Voice quality and naturalness

We will not tell you how OpenAI’s voices sound. No listening test has been run for this review, and describing synthesized speech from documentation would be inventing evidence.

What documentation does settle is the control surface, covered next, and the output formats and caps, covered under API limits.

Expression, pacing, control

This is where OpenAI genuinely differs from the rest of this comparison, and the difference is architectural rather than incremental.

Every other provider here exposes delivery control through markup — SSML tags, or a proprietary tag vocabulary. OpenAI exposes it through prose: the instructions parameter takes natural-language direction about how a line should be delivered. The documentation is explicit about its scope: “Control the voice of your generated audio with additional instructions. Does not work with tts-1 or tts-1-hd.”

The trade is that there is no SSML support anywhere in the speech API. If your pipeline already emits SSML — from a CMS, a localization tool, or an existing integration with another vendor — none of it transfers. Server-sent-event streaming is likewise unavailable on the older models.

What this means for you: instruction steering is excellent for iterating on delivery by hand and poor for deterministic, repeatable output at scale, because prose direction has no formal grammar to validate. Choose it if a human is in the loop tuning the read; think carefully if a machine is generating the direction.

Custom voices

OpenAI is frequently described as a preset-voices-only provider. That framing is out of date, and correcting it is one of the more useful things this review does.

Custom voices exist. The guide states, verbatim: “Custom voices are limited to eligible customers. Contact our sales team to learn more. Once enabled for your organization, you’ll have access to the Voices tab under Audio” (TTS guide, checked September 1, 2026).

The licensing for that feature is referenced but not published. The custom-voices section closes with: “Refer to the Text-to-Speech Supplemental Agreement for additional terms of use.” That sentence carries no hyperlink — we checked the page markup — and the document is not among OpenAI’s published policy list. A buyer therefore cannot read the terms governing custom voices before contacting sales.

We are not going to speculate about what that agreement contains. What we will say is that every other provider in this comparison that offers cloning publishes at least its consent requirements, and here you must ask for them.

API limits and regions

Input is hard-capped at 4,096 characters per request, stated verbatim in the speech endpoint reference: “The maximum length is 4096 characters.” Separately, the gpt-4o-mini-tts model page states a maximum of 2,000 input tokens. Any script longer than that must be chunked client-side, with the seam handling that implies for pacing and prosody across chunk boundaries.

Processing is geographically confined. The speech endpoint’s row in OpenAI’s data-residency table shows storage across all listed regions but processing limited to the United States and Europe (EEA and Switzerland). Regional-processing hostnames exist for a wider set of geographies, but speech is not among the endpoints processed in them. For some buyers that is a simplification; for a buyer with in-country processing requirements outside those two regions, it is a blocker.

One further obligation sits in the content-provenance guidance and is worth quoting because it is unusually direct: “Don’t use repeated queries to reverse-engineer, remove, or evade a watermark.”

Generation speed and latency

We publish no latency figure for OpenAI’s speech models. We measured none, and we did not find a vendor figure in a form we would cite as a specific documented claim.

This section will carry measured figures when our benchmark runs.

Pricing and normalized cost

What you pay, and what you get

On the older models the arithmetic is direct: cost = (characters ÷ 1,000,000) × the rate. A 100,000-character chapter costs $1.50 on tts-1 and $3.00 on tts-1-hd. Those are computations from list prices, not measurements of a workload we ran.

The important limitation: the current model cannot be normalized against the old ones

gpt-4o-mini-tts bills audio output per million tokens, and OpenAI does not publish a token-to-character or token-to-second conversion for audio output. That means the headline comparison a buyer wants — is the new model cheaper than tts-1? — cannot be answered from published information. It also means the new model cannot be compared on price against any character-billed competitor without measuring real usage first.

This is not a rounding problem. It is the single biggest obstacle to evaluating OpenAI’s speech offering on paper, and it is why our per-minute column stays empty here.

Who gets the best value

• Existing OpenAI developers doing modest volumes: strong. One integration, one bill, and the control surface is genuinely better.
• High-volume character-billed work: weak on the old models. $15 per million characters is roughly four times Google’s and Amazon’s standard rates.
• Anyone needing to forecast before building: poor, because the current model’s unit cannot be converted.

Commercial usage, licensing, privacy

The clauses in this section were verified on August 11, 2026. OpenAI’s policy pages returned an access error to every method we tried on September 1, 2026, so they could not be re-read today. They are reproduced with their original verification date rather than presented as freshly checked.

Who owns what you generate

OpenAI’s services agreement contains the most explicit ownership language in this comparison. Clause 4.1, verified August 11, 2026:

“Customer and Customer’s End Users may provide Input and receive Output. As between Customer and OpenAI, to the extent permitted by applicable law, Customer: (a) retains all ownership rights in Input; and (b) owns all Output. OpenAI hereby assigns to Customer all OpenAI’s right, title, and interest, if any, in and to Output.”

An express assignment is stronger than the disclaimers most competitors offer. Read it beside clause 4.4, which carries the same non-uniqueness caveat the whole industry uses: output “may not be unique, and other users may receive similar content”, and responses generated for other users are not your output.

Training

The contractual form of the no-training promise, clause 4.2, verified August 11, 2026: OpenAI “will not use Customer Content to develop or improve the Services, unless Customer explicitly agrees to such use.” The documentation states the same thing in plainer terms, and that page was reachable today: “Your data is your data. As of March 1, 2023, data sent to the OpenAI API is not used to train or improve OpenAI models (unless you explicitly opt in to share data with us)”, with the speech endpoint listed by name in the per-endpoint table.

Set that against the rest of this comparison and it is a real differentiator: Cartesia takes a perpetual, irrevocable licence over inputs and outputs by default, and MiniMax documents no opt-out at all.

A disclosure duty whose source has moved

The text-to-speech guide carries a callout stating, verbatim: “Our usage policies require you to provide a clear disclosure to end users that the TTS voice they are hearing is AI-generated and not a human voice.”

We went to the usage policies to read the requirement in its own words, and did not find it there. The guide asserts an obligation and attributes it to a document that, as far as we can establish, no longer contains it.

Practically, treat the duty as real: it is stated in current official documentation, it costs little to comply with, and a vendor telling you to disclose is not a bad instruction to follow. But if you need to cite the binding source in a compliance document, you will have to ask OpenAI where it now lives.

ChatGPT voices are not API voices

The distinction that catches people is between voice output from ChatGPT and voice output from the API. The service terms have been reported to restrict ChatGPT voice output to non-commercial use and to bar repackaging it as a standalone audio recording. We could not re-verify that clause today because the policy pages were unreachable, so we state the distinction as one to check rather than as a verified quotation.

What is not in doubt is that the two are governed by different documents. If your audio came out of the ChatGPT interface rather than the API, do not assume the API’s ownership assignment covers it.

Pros

• Instruction-steerable delivery in plain language — a control surface no other provider here documents.
• Contractual exclusion of API content from model training, with the speech endpoint named explicitly.
• An express assignment of output rights, the strongest ownership language in this comparison.
• Processing confined to the United States and Europe, simplifying some residency reviews.
• Deprecations published in a dated notice rather than discovered in production.
• One key, one bill and one SDK alongside the rest of the OpenAI platform.
• Prices published as a real table with a Markdown twin that agrees exactly.

Cons

• Two incompatible billing units across three models in one table, with no published conversion.
• No free tier of any kind for speech — the only provider of the ten with none.
• No per-minute cost published for any speech model.
• Custom-voice terms referenced in plain text, unlinked and unpublished.
• A disclosure duty asserted in the guide but not found in the policy it cites.
• 4,096-character hard cap per request, forcing client-side chunking.
• No SSML support anywhere in the speech API.
• Legacy audio families already carry a dated removal notice.
• The tts-1 lineage is described inconsistently across three official pages.
• The older models list at roughly four times the standard rates of the cloud incumbents.

Best use cases

And the inverse, because a recommendation without one is not worth much: if nobody on your team is already using OpenAI, the control surface alone is unlikely to justify the absence of a free tier and the inability to forecast cost before you build.

Alternatives

Organized by the reason you would leave.

• You need a free tier to evaluate. Google Cloud Text-to-Speech gives four million characters a month on Standard and WaveNet (pricing, checked September 1, 2026).
• You need a price you can forecast per character. Amazon Polly lists $4.00 per million characters (pricing, checked September 1, 2026).
• You need self-serve voice cloning. ElevenLabs documents cloning from $6/month (pricing, checked September 1, 2026); Cartesia from $5/month.
• You need SSML. Amazon Polly supports it, with the richest expressive tags on its standard engine.
• You want the whole market in one view. The ranking of ten providers compares pricing, free tiers, licensing and cloning side by side.

Direct comparison links

Head-to-head pages pairing OpenAI against individual competitors are on this site’s roadmap but not published yet, and we do not link to pages that do not exist. Until they are live, the flagship ranking carries side-by-side pricing, licensing and feature tables for all ten providers we cover.

Final verdict

Choose OpenAI when you are already building on the platform and the quality of control matters more than the clarity of the price. Instruction steering is the most interesting idea in this comparison — directing a voice in prose rather than markup removes an entire category of tooling — and the data terms are the best here by a clear margin: a contractual exclusion from training, the speech endpoint named explicitly, and an express assignment of output rights.

Do not choose it when you need to know what it will cost. Two billing units in one table, no per-minute figure, no published token conversion for audio output, and no free tier to measure against together mean a buyer cannot forecast spend before committing engineering time. For a high-volume character-billed workload the older models are several times the price of the cloud incumbents, and the current model cannot be compared to them at all on paper.

What would change this verdict is a conversion table — and, separately, sound. Publishing tokens-per-character for audio output would make this offering comparable overnight. When our audio benchmark runs, this page gains measured comparisons and the assessment table gains scores that mean something. Until then this is a verdict about a price table and a set of terms, one of which we could not re-read today and have dated accordingly.

Frequently asked questions

Sources and what we could not verify

Every changing fact on this page was read from official OpenAI sources on September 1, 2026, except the contract clauses, which carry their own August 11, 2026 date for the reason given above. Prices were cross-checked against two renderings of the same official source, which agree exactly.

What we could not verify

Honesty about gaps beats a page that looks complete. Genuinely unresolved as of September 1, 2026, recorded rather than guessed:

• The contract clauses could not be re-read today. OpenAI’s policy pages returned an access error to every method tried. Clauses are published above with their August 11, 2026 verification date.
• The marketing pricing page could not be cross-checked. It also returned an access error; the developer pricing page was reachable and is the source used.
• Whether ChatGPT voice output is restricted to non-commercial use. Reported in the service terms, not re-verifiable today. Treated as a distinction to check, not a verified quotation.
• Where the AI-disclosure requirement now lives. The guide attributes it to the usage policies; we did not find it there.
• What the Text-to-Speech Supplemental Agreement says. Referenced in plain text without a link and not among the published policies.
• Any token-to-character or token-to-second conversion for audio output. None is published, which is why no per-minute figure appears on this page.
• The current status of tts-1 and tts-1-hd. Three official pages describe their availability differently.
• Any vendor latency figure in a form we would publish.

Change log

September 1, 2026 — First publication. Prices, model status, limits and guide statements read from official OpenAI developer sources on this date; contract clauses carry their August 11, 2026 verification date because the policy pages were unreachable today.

Published September 1, 2026 · The contract clauses will be re-verified and re-dated here when the policy pages are reachable. Benchmark audio, measured costs and scores will be added with a dated entry when the audio benchmark runs (methodology). Independence note: this page contains no affiliate links, and no vendor paid for placement or influenced the order — see how we make money and our editorial policy.