MiniMax Review 2026

Published September 1, 2026 · Every fact below was read from MiniMax’s own developer documentation, pricing tables and platform agreements on September 1, 2026, with the source linked beside the claim. Where a document required an unusual reading method, this review says so.

Quick verdict

Every price, licensing term and feature claim below was read from MiniMax’s official 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:

• Genuinely low pay-as-you-go pricing for long-form synthesis, with a dedicated asynchronous endpoint for long text.
• Clear ownership language on the international platform: you retain your rights in input and generated content.
• Voice cloning and voice design are both available on the standard API rather than gated to enterprise.
• Prices are published as real server-rendered tables, not hidden behind a sales form.
• The mainland synthesis endpoint exposes an explicit watermark parameter, which few competitors offer at all.

Important disadvantages:

• Two platforms, two entities, two governing laws, two currencies and two billing units.
• Cloned and designed voices expire unless used in synthesis within seven days.
• Model training is permitted by default and no opt-out is documented anywhere.
• The international subscription’s billing unit, “audio points”, is never defined in the documentation.
• No speech-to-text model exists, so voice agents need a second vendor.
• No MiniMax-branded speech SDK in any language.
• Two model identifiers are accepted by every speech endpoint but documented and priced nowhere.
• The international endpoint has no watermark parameter, while the customer carries a disclosure duty.

On this page

Two platforms, two legal regimes

This comes first because almost every other fact on the page depends on which platform you are actually buying from.

What this means for you: a price comparison between the two properties is not a like-for-like comparison, and neither is a legal review. If your procurement team approved the international terms, the mainland platform is a different contract with a different forum and a different remedy path. Choose deliberately, and record which one you are on.

Who should use it — and who should look elsewhere

Choose MiniMax if:

• Long-form synthesis at low cost is the requirement. Pay-as-you-go pricing is published openly, and a dedicated asynchronous long-text endpoint exists for exactly this shape of work (pricing, checked September 1, 2026).
• You want cloning and voice design on the standard API. Both are available without an enterprise conversation, which is not true of several competitors here.
• You can accept Singapore-law arbitration. The international terms are clear about the forum; the question is whether your organisation can live with it.
• You need a watermark parameter on the request. The mainland synthesis endpoint exposes one — unusual in this market — though the international one does not.

Look elsewhere if:

• You need speech-to-text from the same vendor. MiniMax publishes no ASR model at all, so a voice agent needs a second supplier. Cartesia sells both on one account (pricing, checked September 1, 2026).
• You cannot have your inputs and outputs used for model development. It is permitted by default and no opt-out is documented on either property.
• You want a native SDK. There is no MiniMax-branded speech SDK in any language.
• You need to budget a subscription precisely. The international audio subscription is denominated in “audio points”, a unit the documentation never defines.

Current price, free plan, commercial rights

The short version: pay-as-you-go pricing is published and cheap, the subscription unit is undefined, and there is no documented free speech tier.

Read from the “Audio” section of the international pay-as-you-go page and the 语音 section of the mainland equivalent, both checked September 1, 2026. Both are server-rendered tables rather than JavaScript-injected values. We publish both units as the vendor states them and perform no currency conversion; converting would produce a number MiniMax does not publish.

The same asynchronous long-text endpoint is priced at the same rates on both properties for the respective models.

The subscription unit is never defined

The international platform sells an audio subscription whose tiers are denominated in “Audio points per month” — 100,000, 300,000, 1,100,000, 3,300,000 and 20,000,000. What an audio point is is not stated. The phrase occurs exactly once across the complete international documentation index we read, and never in a definition.

What this means for you: you cannot compute the effective per-character cost of any subscription tier from published information. Until MiniMax defines the unit, the pay-as-you-go rates are the only figures a buyer can actually budget from — and we would build a projection on those, not on points.

One documented exclusion is worth knowing before you buy a plan for cloning work: the token plan’s own footnote states that coverage spans the model lineup including speech, but that a small number of special models are excluded — and voice design and rapid voice cloning fall outside it. A subscription is not a route to cheap cloning here.

Commercial rights and ownership

The international terms, effective March 30, 2026, address ownership and training in a single clause of the intellectual-property section. Verbatim:

“As between you and us, and to the extent permitted by applicable laws, you retain your ownership rights in Client input and generated content. We may use the input and generated content to provide, maintain, develop, and improve our Services, comply with applicable law, enforce our terms and policies, and keep our Services safe.”

Read the two sentences together, because they are doing different jobs. The first grants you ownership. The second reserves a broad use right over the same material — including to “develop” and “improve” the services — and it is not qualified by anonymisation, by plan, or by anything else. The mainland document words ownership conditionally instead, making it contingent on what local law provides.

We found no opt-out from that use on either property. The phrase does not appear in the documents we read.

What this review covers

This review covers MiniMax’s speech synthesis offering on both the international and mainland platforms: the models and their identifiers, the published pay-as-you-go prices in their own units, the subscription structure, cloning and voice design including their lifecycle, the rate limits, and the terms governing ownership, training, disclosure and disputes.

It does not evaluate how any MiniMax 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 note on how the legal documents were read, because it affects how much weight to give them. Unlike the pricing tables, which are ordinary server-rendered HTML, MiniMax’s platform agreement pages are client-side rendered applications: the served HTML contains almost no visible text, and the full agreement sits inside the page’s data payload. We read the agreement text out of that payload rather than from a rendered browser view. It is unambiguously the vendor’s own served document, and a person opening the URL in a browser sees the same text — but we would rather tell you how we obtained it than imply we read it the ordinary way.

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: prices in both currencies and units, model enums on every speech endpoint, cloning lifecycle and gating, rate limits on both properties, and the contract clauses quoted below, all read on September 1, 2026.

Assessment by dimension

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

Models, including two that are not documented

The current speech line is speech-2.8-hd and speech-2.8-turbo, priced on both properties. Earlier generations — speech-2.6-hd, speech-2.6-turbo, speech-02-hd and speech-02-turbo — sit inside a collapsed documentation block titled “Legacy Models” on the international property, and its Chinese equivalent on the mainland one, while remaining priced on the current pay-as-you-go page.

No dated end-of-life exists for any speech model on either property. That is a genuine and useful contrast with several competitors in this comparison, and we checked for it specifically. Note that an adjacent audio product — the music and lyrics APIs — does carry a dated service-adjustment notice, but it does not apply to speech.

Two identifiers deserve a warning. speech-01-hd and speech-01-turbo are accepted values in the model enum of every speech endpoint on both properties — synchronous, WebSocket, bidirectional, asynchronous and voice cloning — yet they appear in no models table and carry no published price. An API that accepts a model it does not document is an API that can bill you for something you cannot look up. Do not send either identifier without confirming what it costs.

Voice quality and naturalness

We will not tell you how MiniMax sounds. No listening test has been run for this review.

Nor is there a vendor figure to quote: we found no measured quality or latency number for any MiniMax speech model on either property, across every speech guide, both API references, the API overview, both release-notes pages, the audio FAQ and both pricing pages. That is unusual — most vendors publish at least a marketing benchmark — and it means this section has nothing in it until our own benchmark runs.

Voice cloning, consent, and the seven-day timer

Cloning and voice design are both available on the standard API, which is a real advantage over providers that gate cloning behind enterprise plans. Two documented behaviours change how you should use them.

Voices are temporary until used. The API overview states verbatim: “Voices produced by cloning and voice design are temporary: the fee is charged only on first use in speech synthesis (previews within those APIs do not count)”. The practical effect is a timer: create a clone, fail to use it in synthesis, and it goes away. If your workflow clones in a batch and synthesizes later, that gap is where you lose voices.

Consent is not enforced by the product. Cloning is identity-gated on the developer — the account holder must be verified — but we found no voice-owner consent requirement published anywhere in the developer documentation. The cloning flow takes an uploaded audio file and returns an identifier; it does not ask you to attest that the speaker agreed.

That places MiniMax at the permissive end of this market. Speechify’s API, by contrast, requires the speaker to record a phrase the vendor issues, with no unattended path (consent docs, checked September 1, 2026). If your compliance position depends on the platform enforcing consent, it does not here — and note that you also carry a disclosure duty, described in licensing.

API, SDKs, rate limits

There is no MiniMax speech SDK. The platform’s only SDK page is titled “Integrate via SDK” and its subtitle directs developers to use a third party’s SDK to call a MiniMax language model — not a speech one. Speech integration is direct HTTP.

There is no speech-to-text model. The complete international documentation index contains no ASR, transcription or speech-to-text page and no such model. If you are building a voice agent, MiniMax can speak but cannot listen, and you will need a second vendor for the other half.

Published rate limits for speech on the international property are 60 requests per minute for text-to-audio, 60 for voice cloning and 20 for voice design. The mainland ceiling is documented lower. We found the rate-limit documentation internally inconsistent in places and do not attempt to reconcile it here; treat the published figures as the floor of what to design for and confirm against your own account.

Generation speed and latency

No latency figure of any kind is published for any MiniMax speech model on either property, and we have measured none. This section has nothing to report — not even a vendor claim to quote — and will carry measured figures when our benchmark runs.

Pricing and normalized cost

What you pay, and what you get

On the international platform the rate is the effective cost, because billing is per character with no subscription required: cost = (characters ÷ 1,000,000) × the model rate. A 100,000-character audiobook chapter comes to $6.00 on speech-2.8-turbo and $10.00 on speech-2.8-hd, at the prices published on September 1, 2026. That is arithmetic from list prices, not a measurement of a workload we ran.

The important limitation: the subscription cannot be normalized at all

Because “audio points” is undefined, there is no way to compute what a subscription tier costs per character, and therefore no way to know whether it beats pay-as-you-go. That is not a gap in our research — we looked across the complete documentation index — it is a gap in the published information. Budget on the pay-as-you-go rates.

Who gets the best value

• Long-form batch synthesis: strong value on the turbo model, with a purpose-built asynchronous endpoint.
• Cloning-heavy workflows: mixed. Cloning is available on the standard API, but excluded from the token plan and subject to the expiry timer.
• Voice agents: poor fit at any price, because there is no speech-to-text half.

Commercial usage, licensing, privacy

Ownership, and the use right beside it

The international clause is quoted in full in commercial rights above. Its structure is worth restating plainly: you retain ownership of input and generated content, and MiniMax reserves a broad right to use that same content to develop and improve its services. Both halves are in one clause and there is no separate output-licence clause anywhere in the agreement.

The mainland document words ownership conditionally — contingent on what applicable law provides — rather than granting it outright. Same company, two different formulations, and which one applies depends on which platform you signed up to.

The same intellectual-property section places warranties on you regarding the inputs you supply, with an indemnity tail. If you are cloning a voice, that is the clause that allocates the risk of having got consent, and it allocates it to you.

You carry an AI-disclosure duty

MiniMax’s international terms impose an affirmative obligation on the customer concerning the labelling of AI-generated content. This is not a courtesy request buried in a policy; it sits in the user-obligations section of the agreement effective March 30, 2026, and it applies on the Singapore-entity international terms, not only on the mainland ones.

Set against that, the international synthesis endpoint exposes no watermark parameter at all. We enumerated every property of the international text-to-audio request object — model, text, stream, stream options, voice settings, audio settings, pronunciation and the rest — and none of them controls a watermark. The mainland endpoint does expose one.

So the obligation to label is yours, and on the international platform the tooling to do it in the audio itself is not provided by the API. Plan your disclosure mechanism outside the request.

Disputes and data location

The international terms select Singapore law with binding SIAC arbitration — no court option and no small-claims carve-out — and state that personal data is “stored in the data center located in the United States”. The mainland terms select PRC law and litigation. For most buyers outside China the international regime is the relevant one, and the arbitration clause is the sentence to take to counsel before signing.

Pros

• Low published pay-as-you-go pricing for long-form synthesis, in real server-rendered tables on both properties.
• A dedicated asynchronous endpoint for long text, priced the same as the synchronous one.
• Voice cloning and voice design on the standard API, without an enterprise gate.
• Clear ownership grant on the international platform: you retain rights in input and generated content.
• No dated end-of-life for any speech model on either property — a real contrast with several competitors here.
• The mainland synthesis endpoint exposes an explicit watermark parameter.
• Identity gating on cloning developers is at least a deliberate abuse control.

Cons

• Two platforms with different entities, governing law, dispute mechanisms, currencies and billing units.
• Cloned and designed voices expire unless used in synthesis.
• Broad default use right over inputs and outputs, with no documented opt-out on either property.
• The subscription billing unit, “audio points”, is never defined.
• Voice design and rapid voice cloning are excluded from the token plan.
• No speech-to-text model, so voice agents need a second vendor.
• No MiniMax-branded speech SDK in any language.
• Two model identifiers accepted by every speech endpoint are documented and priced nowhere.
• A customer disclosure duty with no watermark parameter on the international endpoint.
• Binding arbitration with no court option on the international terms.

Best use cases

And the inverse, because a recommendation without one is not worth much: if your organisation needs to point at a single contract, in a single currency, with a defined billing unit and an opt-out from model training, MiniMax does not offer that on either of its platforms today.

Alternatives

Organized by the reason you would leave.

• You need speech-to-text from the same vendor. Cartesia sells synthesis and transcription on one account (pricing, checked September 1, 2026).
• You need cloning with enforced consent. Speechify’s API requires a recorded consent phrase with no unattended path (consent docs, checked September 1, 2026).
• You need a defined billing unit and a single jurisdiction. Amazon Polly bills per character with ownership in a numbered clause (pricing, checked September 1, 2026).
• You want to avoid a default training licence entirely. Resemble AI publishes MIT-licensed models you can self-host with no vendor account.
• 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 MiniMax 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 MiniMax when the job is long-form synthesis, the budget is tight, and you are comfortable with the international terms. The pay-as-you-go rates are genuinely low, the asynchronous long-text endpoint is the right tool for audiobook-shaped work, cloning is available without a sales call, and — unusually in this comparison — no speech model carries a published end-of-life date hanging over it.

Do not choose it when your project needs the listening half of a voice agent, an opt-out from model training, a subscription you can price, or a single legal regime. The two-platform split is not cosmetic: different entities, different law, different dispute mechanism, different currency, different billing unit, and ownership granted flatly in one document and conditionally in the other. Add a cloning lifecycle that quietly expires unused voices and a disclosure duty with no watermark parameter to satisfy it, and the low headline price is buying you a set of obligations to manage.

What would change this verdict is sound — and, before that, a definition. Publishing what an audio point is would make the subscription comparable overnight; documenting the two orphaned model identifiers would close a billing risk. 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 two rate cards and two contracts.

Frequently asked questions

Sources and what we could not verify

Every changing fact on this page was read from official MiniMax sources on September 1, 2026. Prices come from server-rendered tables. Agreement text was read from the served page payload, as explained in what this review covers. Where the two properties differ, both are published and neither is treated as the default.

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:

• What an “audio point” is. The term is used to price the audio subscription and defined nowhere in the documentation we read.
• What speech-01-hd and speech-01-turbo cost or do. Both are accepted by every speech endpoint’s model enum and appear in no models table and no price list.
• The correct rate limits. The published rate-limit documentation is internally inconsistent; we report the figures as published without reconciling them.
• Any quality or latency figure for any speech model on either property. None exists to quote, even as a vendor claim.
• Whether an opt-out from model-development use exists. We found none documented on either property; we record its absence from the pages read rather than asserting none exists anywhere.
• Whether the mainland ownership wording produces the same result as the international wording. One grants ownership flatly, the other conditions it on applicable law. That is a question for counsel, not for us.
• Whether the international platform offers any watermarking mechanism outside the request object. The endpoint exposes no parameter; we did not find an alternative mechanism documented.
• Any free speech tier. None is documented on the pages we read.

Change log

September 1, 2026 — First publication. All prices, terms, model identifiers, limits and feature statements read from official MiniMax sources on this date, across both the international and mainland platforms.

Published September 1, 2026 · Benchmark audio, measured costs and scores will be added with a dated entry here 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.