Home › Rankings

Best AI voice generators of 2026: you are not comparing voices, you are comparing three ways to be billed

Data last updated: 2026-10-14. Pricing may have changed — verify with the vendor.

AITop is an independent guide. Prices and features change fast — always check the vendor's page.

New to the jargon? AITop's AI glossary explains "voice cloning", "character" and the rest in plain English.

Related: what voice cloning is · commercial-use rights · watermarking & provenance · the best AI transcription tools

The 30-second answer

Three vendors sell you credits on a monthly plan, two sell you characters metered like cloud bandwidth, and one sells you tokens. That shape decides your bill far more than which voice sounds best. ElevenLabs has the best-sounding output and the worst price per million characters; Amazon Polly and Google Cloud are roughly 45× cheaper per character and include commercial rights in the meter; Cartesia is the one built for real-time agents; OpenAI is the least friction if you already call its API. And the free tier that looks most generous — ElevenLabs, PlayHT, Cartesia — is the one whose licence normally forbids publishing.

1

ElevenLabs9.1

The best speech you can buy, and the most expensive per character by an order of magnitude — both statements are true, and the second one is the whole decision. The plan structure is a monthly credit pool shared across every product: Free $0 (10,000 credits), Starter $6 (30,000), Creator $22 (121,000), Pro $99 (600,000), Scale $299 (1,800,000, three seats), Business $990 (6,000,000, ten seats). For the V2 multilingual models 1 text character equals 1 credit, so Creator works out to roughly $182 per million characters and Business to about $165 — there is no tier that brings it near the mid-market. Three things matter more than the price. First, the licence: commercial use starts at Starter, and the free tier does not include it. Second, cloning is tiered — Instant Voice Cloning from Starter, Professional Voice Cloning from Creator. Third, the meter is shared, so the same credits pay for text-to-speech, speech-to-text (330 credits per minute), music (900 credits per minute) and sound effects; a project that uses several of those can exhaust a plan you sized for narration alone.

Pricing: Free $0 (10k credits) · Starter $6 (30k) · Creator $22 (121k) · Pro $99 (600k) · Scale $299 (1.8M, 3 seats) · Business $990 (6M, 10 seats) · 1 credit per character on V2 multilingual · effective $182 per million characters at CreatorBest for: Anyone whose output is judged by ear — audiobooks, ads, character work
2

Cartesia (Sonic 3.5)8.7

The one vendor here that is built for a person waiting for a reply rather than a listener enjoying a file. Cartesia bills one credit per character, flat, on every Sonic generation: Free $0 (20,000 credits a month), Pro $5 (100,000), Startup $49 (1,250,000), Scale $299 (8,000,000). At full use that is $50 per million characters on Pro, $39.20 on Startup and $37.38 on Scale — the best volume economics in this list after Amazon, with a commercial licence from Pro and instant cloning included. Two details decide whether it is right for you. Credits roll over, capped at twice your monthly allowance, which is worth more than a small rate difference if your workload is lumpy. And the overage setting defaults to off, which means requests are blocked at zero balance rather than billed — on a production endpoint that is a hard stop, not a rounding error. The real-time work is metered separately again: managed agent calls add roughly $0.06 a minute on top, and the credits are not the whole invoice.

Pricing: Free $0 (20k credits/mo) · Pro $5 (100k) · Startup $49 (1.25M) · Scale $299 (8M) · 1 credit per character · effective $50 / $39.20 / $37.38 per million chars · agent calls ~$0.06 per minuteBest for: Voice agents, phone bots and anything conversational that cannot wait
3

OpenAI TTS8.3

Not the best voice, not the cheapest, and frequently the right answer anyway — it is one more endpoint on a key you already have, inside a bill that already gets approved. Two flat-rate models price by character: tts-1 at $15 per million characters and tts-1-hd at $30. The newer recommended model, gpt-4o-mini-tts, prices by token instead — $0.60 per million text input tokens plus $12 per million audio output tokens, which works out to roughly $0.015 per minute of audio, and it steers delivery through a natural-language prompt rather than speed dials. Read the three limits before you commit. There is no free tier for TTS at all, so it is the only vendor here you cannot try without a card. Custom voices are gated behind a sales conversation rather than a self-serve upload. And mini-tts caps at 2,000 input tokens per request, so a chapter means chunking the script yourself and stitching the audio — which is exactly the work ElevenLabs' Studio and PlayHT's workbooks exist to remove.

Pricing: tts-1 $15 per 1M characters · tts-1-hd $30 per 1M characters · gpt-4o-mini-tts $0.60 per 1M text input tokens + $12 per 1M audio output tokens (~$0.015 per minute) · no free tier · commercial use includedBest for: Developers already calling OpenAI who need speech in an afternoon, not a voice brand
4

Google Cloud Text-to-Speech8.0

The cheapest published character rate in this list, with the most confusing price list inside one product — the spread between its own tiers is 40×. Standard and WaveNet voices are $4 per million characters with the first 4 million free every month and no 12-month cliff; Neural2 and Polyglot are $16 per million with 1 million free; Chirp 3: HD is $30 per million with 1 million free; Instant custom voice is $60 per million; and Studio voices sit at $160 per million, which is the same order of magnitude as a premium creator subscription. Commercial use is included in the usage price, and if your stack already lives in Google Cloud there is no new vendor to paper. What you do not get is a product: this is an API and a console, with no timeline editor, no dubbing studio and no pronunciation dictionary workflow. The other trap is naming — "Google TTS" is too broad to quote at a single price, because the model family is the bill.

Pricing: Standard / WaveNet $4 per 1M characters (4M free per month, ongoing) · Neural2 $16 (1M free) · Chirp 3: HD $30 (1M free) · Instant custom voice $60 · Studio $160 (1M free) · commercial use included · model family determines the rateBest for: High-volume narration, accessibility and IVR where the rate matters more than the timbre
5

PlayHT7.7

The cheapest way into a real creator subscription, and the one that behaves most like a workbench. Starter is free forever with 10,000 credits a month, one voice-clone slot and MP3 out. Creator is $9.99 a month ($99.99 a year) with a million credits, three clone slots, MP3 and WAV, five workbooks and a commercial licence. Studio is $34.99 a month ($349.90 a year) with 3.5 million credits, ten clone slots, ten seats, unlimited workbooks and — this is the catch — API access, which Creator does not include. Scale is quote-only with SSO, a DPA and a named contact. Two things to know going in. The free Starter tier is for personal use, not publishing. And at larger volumes the effective rate stops being the story: Creator's million credits for $9.99 works out around $10 per million characters, which is genuinely competitive, while the step up to API access forces the $34.99 tier whether or not you need its seats.

Pricing: Starter $0 (10k credits, personal use) · Creator $9.99/mo or $99.99/yr (1M credits, commercial licence, 3 clone slots) · Studio $34.99/mo or $349.90/yr (3.5M credits, 10 seats, API access) · Scale quote-only · credits roll over on paid plansBest for: Freelancers and small teams who want a workbench plus a licence for under $10 a month
6

Amazon Polly7.4

The floor of the market. Standard voices are $4 per million characters with 5 million characters free every month — the most generous ongoing allowance of any vendor here — Neural voices are $16 per million with 1 million free for the first 12 months, Generative is $30 with 100,000 free for 12 months, and Long-Form is $100 with 500,000 free for 12 months. Commercial use is included in the usage price, and cached speech can be replayed at no extra cost, which matters if you are rendering the same prompts repeatedly. It ranks last for the same reason it is cheap: there is no creator product at all. No studio, no workbooks, no dubbing, no voice marketplace — just an API and a set of voices, plus the operational debt of an AWS account. If your audio is transactional (an IVR reading account balances, a notification read aloud, a preview in a CMS), this is the correct answer and everything above it is overspend. If a human will sit and judge the voice, it is not.

Pricing: Standard $4 per 1M characters (5M chars/month free, ongoing) · Neural $16 per 1M (1M/month free for 12 months) · Generative $30 per 1M (100k/month free, 12 months) · Long-Form $100 per 1M (500k/month free, 12 months) · commercial use includedBest for: Transactional speech at volume where nobody is admiring the prosody

What the comparison tables hide

Every free tier hides the same sentence: not for commercial use

Read the licence line before the credit count, because the two are unrelated. ElevenLabs withholds commercial rights on Free and grants them from Starter at $6. PlayHT's free Starter tier is explicitly personal use. Cartesia's commercial licence starts on Pro. OpenAI publishes no free tier for text-to-speech at all, which at least removes the ambiguity. The metered cloud APIs are the exception: Google Cloud and Amazon Polly fold commercial use into the per-character price, so the "expensive" option is often the one you can actually ship with. A free plan is a demo of the voice, not a licence to publish it.

The unit you are quoted in is not the unit you publish

Characters, credits, minutes and tokens are four different rulers, and two vendors in this list bill one product in two of them. ElevenLabs charges one credit per character on the V2 multilingual models. Cartesia charges one credit per character. PlayHT sells credits. OpenAI's tts-1 bills per character while its recommended gpt-4o-mini-tts bills per token — input text tokens plus audio output tokens — and because it publishes no characters-per-token ratio, the two cannot be placed on the same axis without testing your own script. Every vendor here is willing to tell you the price of its unit. None of them will tell you how many units your script is, and that is the number your budget actually needs.

Real-time is a different product on a different meter

Conversational speech and narration look like the same feature and are priced as separate businesses. Cartesia's managed agents add roughly $0.06 per call minute on top of its character credits, plus telephony on top of that. OpenAI separates the realtime models from plain TTS entirely, and they cost multiples of the TTS rate because every turn is billed in both directions while a human is waiting. The common and expensive mistake is buying a narration plan to power a phone agent, or paying real-time prices for a podcast you render once and publish. Decide which of the two you are building before you compare a single number.

Certified price data is a promise the vendor is not making

Every figure on this page was read from the vendor's own pricing page and cross-checked against at least one independent source on the same date, which is exactly why three well-known tools are missing from it. Where a vendor's published rates disagreed between its own pages, or where a tier's effective price derives from a promotional discount with an end date, we left the product out rather than rank a number we cannot reproduce. That makes this list shorter than most. It also makes each row something you can go and check, which is the only useful property a price table has.

Straight answers

Per-character or per-credit — which one should I compare?

Neither, directly. Convert both to the only number that means anything: cost per finished minute of audio you can actually publish. Take your own script, count its characters, and run it through each vendor's model — then divide by the duration you got. Two things fall out. The first is that ElevenLabs' effective rate is a multiple of the cloud APIs, not a percentage more, and only you can say whether the difference is audible in your use case. The second is that a cheap rate on a plan you cannot publish under is not cheap, it is a trial, and the trial ends the moment you want to ship.

Can I publish audio made on a free plan?

Usually not, and the answer is written in the licence rather than the pricing table. ElevenLabs: no commercial rights on Free, yes from Starter. PlayHT: the free Starter tier is personal use. Cartesia: commercial licence from Pro. OpenAI: no free tier exists, so there is nothing to misinterpret. Google Cloud and Amazon Polly: commercial use is part of the metered price. If you are publishing — a course, an ad, a product — treat the free tiers as audition reels and budget a licence from day one, because re-recording a catalogue later costs more than the plan ever did.

Why aren't Speechify, Murf, Azure AI Speech or WellSaid Labs on this list?

Because we could not confirm their current list prices from two independent sources on the same date, and one of them no longer publishes a public price at all. That is the same bar we apply everywhere on this site: a vendor whose own pages disagree about its own tiers does not get a row, because the number would be decoration rather than data. Every vendor here is ranked on prices we could read twice and reproduce, which is why the list is six long and not ten. If you use one of the missing tools and can point at its current public pricing page, we will review it and, if it holds up, rank it.

The verdict

Pick the billing shape that matches your volume, then check the licence — in that order. If a human will listen and judge the output, ElevenLabs is still the ceiling and the $22 Creator plan is where the licence starts, not the free tier. If you are building a voice agent, Cartesia is the only vendor here priced for a person waiting, and the rollover and the blocked-at-zero overage are the two settings to change on day one. If you already call OpenAI, nothing beats adding speech to a key you have, provided you can live without self-serve cloning and chunk your own long scripts. If your audio is transactional, Google Cloud at $4 per million with 4 million free every month or Amazon Polly at $4 per million with 5 million free is the honest answer, and paying 40× more for prosody nobody hears is a branding decision rather than an engineering one. And whichever you choose, read the commercial-use line before the credit count: that sentence, not the rate, decides whether you can ship it.