What are the best ElevenLabs alternatives if per-character billing is the problem?
Sort them by how they bill, not by how many voices they list. There are three kinds. Cheaper metered services cut the price of a character. Open models remove the meter, if you can install and run them. A local app removes it on a Mac for a single payment. Which kind fixes your problem depends on whether the rate hurts or the meter does.
Most lists of ElevenLabs alternatives rank cloud services against each other on voice count, language count and latency. Those are fair questions for a team building a voice agent. They are the wrong questions for someone narrating videos, articles or chapters every week, whose bill grows with every minute they publish and every take they throw away.
Three kinds of alternative
A cheaper meter. Fish Audio, Cartesia and OpenAI's speech API all charge less per character than ElevenLabs does. You still pay for every render, and the plan still decides how much you can make in a month. This is the right move when the rate is the problem and your volume is modest.
No meter, your own setup. Kokoro, Chatterbox and VoxCPM2 are open models with permissive licences. You download the weights and run them on your own hardware. Nothing counts characters, but you own the installation, the Python environment and the updates.
No meter, packaged. txt2mp3 puts VoxCPM2 inside a Mac app sold for one payment. The model runs on the Mac's own silicon, so there is no per-character price after you buy it. It only runs on Apple silicon, which rules it out for anyone on Windows or Linux before price comes into it.
Why the rate is not the whole bill
A per-million rate is the price of one take of a script, sent once, on a plan you keep forever. Real narration breaks all three assumptions. A cheaper meter fixes the rate. The other three costs stay, and they decide whether switching changes much.
Retakes
Every metered service bills the text you send, every time you send it. A name comes out wrong, a line lands flat, you rewrite a sentence once you hear it read. Two takes per section is a modest habit for anyone who cares how the read sounds, and it doubles the generated minutes behind every published one. At three takes the bill triples. None of the alternatives below discounts a second take; it costs what the first did. We work through the arithmetic in what per-character billing actually means.
Exit cost
The second cost arrives when you stop paying. ElevenLabs' pricing page states that downgrading or cancelling forfeits unused credits at the end of the cycle. Its rollover, up to two months and at most three times the monthly quota, exists only while a paid plan is active. Prepaid credit is a feature of the subscription, not a balance you own.
A cloned voice is the same kind of asset. On a cloud service it lives in the account. Play.ht is the cautionary case: Meta acquired the company in July 2025 and the team joined Meta. Per reporting on the shutdown, the platform closed on 31 December 2025 and accounts, voice clones, saved audio and API keys were deleted with no export tool. A channel whose narrator lived there lost the narrator, whatever plan it was on. What each service keeps and locks when you simply stop paying is set out in what happens to your voice when you cancel a TTS subscription.
Cheaper meters: Fish Audio, Cartesia and OpenAI
These three charge noticeably less per character than ElevenLabs, and two of them clone voices. Fish Audio's API costs $15 per million bytes of text, Cartesia's $5 plan buys 100,000 characters, and OpenAI's tts-1 is $15 per million characters. All three still meter every render, so a retake and a lapsed plan cost what they always did.
For scale: ElevenLabs' own API charges $0.08 per 1,000 characters on its multilingual models and $0.04 on Flash and Turbo, the same on every tier. Inside a subscription, the Creator plan is $22 a month for 121,000 credits, which ElevenLabs lists as about 121 minutes, with extra minutes at about $0.18.
Fish Audio
The closest like-for-like swap for a creator. The Plus plan is $11 a month for 250,000 credits, which Fish Audio lists as up to 200 minutes of generation, with commercial use allowed and voice cloning included. Its FAQ puts a minute at roughly 600 to 625 credits. Pro is $75 for 2,000,000 credits, about 1,620 minutes.
Two details matter. The free plan's 8,000 credits, about seven minutes, are for personal, non-commercial projects only, according to the FAQ, even though the plan card lists commercial use. And the API bills UTF-8 bytes rather than characters: $15 per million for the s2-pro and s2.1-pro models. English text is one byte per character. Accented letters take two, so a German or French script costs more than its character count suggests.
Cartesia
Built for low-latency voice agents, and cheap to start. Pro is $5 a month for 100,000 credits, with a commercial licence and instant voice cloning; one credit is one character. Startup is $49 for 1.25 million credits and Scale is $299 for 8 million. Professional voice clones start on Startup. The free plan's 20,000 credits list no commercial licence.
If you are building something that has to answer a caller in real time, Cartesia is a better fit than anything else on this page. For batch narration the speed advantage matters less, and you are paying for a feature the job does not use.
OpenAI text to speech
The lowest per-character rate on this page, tied with Fish Audio's API, and an API rather than an app. tts-1 costs $15 per million characters and tts-1-hd $30. There are 13 built-in voices. Custom voices exist, but only as an approved voice created from a speaker's consent recording, which is a different thing from uploading a clip. OpenAI's usage policies also require you to tell listeners clearly that the voice is AI-generated.
For a developer who wants speech inside their own software, this is hard to beat on price. For a narrator it means writing code or finding a front end, and the front end usually adds its own margin on top of the rate.
Murf is the other name that comes up on every list. It gets its own page in Murf vs txt2mp3, so we leave it out of the arithmetic here.
ElevenLabs alternatives with voice cloning
Cloning is common now; owning the clone is not. Fish Audio and Cartesia both clone, Cartesia from its $5 plan, and OpenAI only through an approved consent recording. In each case the voice lives in the account, not on your disk. Open models and a local app clone from a recording on your disk, and the voice stays a file you keep.
ElevenLabs itself offers instant cloning from Starter and professional cloning from Creator. So switching to a cheaper cloud service does not change where your voice lives. It changes which company holds it.
The useful question before you clone anywhere is the one Play.ht customers learned to ask too late: if this account closed tomorrow, what would I still have? On a cloud service, the answer is your reference recording and whatever audio you already exported. On your own machine it is the voice itself, because the voice was never anywhere else.
Open models you run yourself
Kokoro, Chatterbox and VoxCPM2 are free to download under permissive licences, and they remove the meter entirely. The cost moves to setup: a Python environment, the model weights, a GPU or an Apple silicon Mac, and your own time when something breaks. For a tinkerer that is an afternoon. For a narrator it is a side project.
- Kokoro is an 82-million-parameter model under Apache 2.0. Its v1.0 release lists 8 languages and 54 voices. The model card does not mention cloning, so it is a preset-voice engine.
- Chatterbox, from Resemble AI, is MIT-licensed. The Turbo model has 350 million parameters
and is English-only; the multilingual model has 500 million and covers 23 or more languages.
It clones from a reference clip, and its examples run on CUDA, CPU or Apple's
mpsdevice, with a Mac example in the repository. - VoxCPM2, from OpenBMB, is a two-billion-parameter model under Apache-2.0, free for commercial use. It covers 30 languages, outputs 48 kHz audio, clones from a short clip and can design a new voice from a written description. Its authors report state-of-the-art or competitive results on the major zero-shot and controllable benchmarks. The model card quotes about 8 GB of VRAM on NVIDIA hardware.
None of these is a product. There is no editor, no history, no export button and no one to ask. Long scripts have to be split, the pieces stitched back together and the takes tracked by hand. That work is the real price of free. Our guide to running text-to-speech locally on a Mac walks through the do-it-yourself route step by step.
An offline ElevenLabs alternative on a Mac
txt2mp3 packages VoxCPM2 into a Mac app sold for one payment: $29 for the launch batch, then $69, then $99 as the standing price. It clones and designs voices, renders with the network off, and has no character meter after purchase. The trade is the platform: it runs on Apple silicon Macs and nowhere else.
Setup downloads about 5 GB once, the speech model and the Python runtime that drives it. After that, every take renders on the Mac's own silicon. On our development Mac, loading the model takes 18.4 seconds and a short take renders in 5.5 seconds; the app keeps the model loaded, so only the first take of a session pays for the load.
What one payment covers
Three routes to a narrator: 14 preset voices, a clone made from a recording you own, or a voice described in words. All three are files in your library. Exports are MP3 at 192 kbps or WAV, with word-level caption timings aligned on the same machine. The commercial right to use what you generate is part of the purchase, and it does not lapse, because nothing is rented.
The second take is free, and so is the ninth. That is the change that alters how you work rather than what you spend: a retake stops being a budget decision. A side effect of the same design is that the script never leaves the machine, which matters for client work and unreleased episodes.
The trial speaks 120 characters before you pay anything, so you can judge the voice on your own text. A few other local Mac voice apps exist; we compare them in the best offline text-to-speech apps for Mac.
What a finished hour costs on each
ElevenLabs lists 121,000 credits as about 121 minutes, so a spoken minute is roughly 1,000 characters and an hour is about 60,000. One take of that hour costs $0.90 on OpenAI tts-1 or Fish Audio's API, $3 on Cartesia Pro, $4.80 on ElevenLabs' API and about $11 inside the Creator plan. Two takes double every figure; a local app stays at zero.
| Service | Rate used | One take of an hour | Two takes | Three takes |
|---|---|---|---|---|
| OpenAI tts-1 | $15 per 1M characters | $0.90 | $1.80 | $2.70 |
| Fish Audio API (s2-pro) | $15 per 1M bytes | $0.90 | $1.80 | $2.70 |
| OpenAI tts-1-hd | $30 per 1M characters | $1.80 | $3.60 | $5.40 |
| Cartesia Pro | $5 for 100,000 credits | $3.00 | $6.00 | $9.00 |
| Fish Audio Plus | $11 for up to 200 minutes | $3.30 | $6.60 | $9.90 |
| ElevenLabs API, multilingual | $0.08 per 1,000 characters | $4.80 | $9.60 | $14.40 |
| ElevenLabs Creator | $22 for about 121 minutes | about $11 | about $22 | about $33 |
| txt2mp3 | One payment | $0 | $0 | $0 |
The Fish Audio API row assumes English, at one byte per character. The plan rows assume you use the whole allowance; an unused month raises the real rate.
The table says something the rankings tend to skip. At pure API rates, a metered service is very cheap. $29 buys about 32 hours of single-take generation on OpenAI tts-1. If that is your volume and you are happy with a preset voice and some code, the purchase does not pay for itself quickly, and we would not pretend otherwise.
Break-even against each
The purchase wins against plans, not against raw APIs at low volume. Against ElevenLabs Creator at $22 a month, the $29 launch price is repaid in month two and the $99 standing price in month five. Against Fish Audio Plus at $11, those become month three and month nine. Against Cartesia Pro at $5, month six and month twenty. The full ElevenLabs timeline, with annual billing and the first-month discount, is in ElevenLabs vs txt2mp3.
The direction only goes one way. More output, more takes or a cloned voice all push you up a plan tier, which moves the crossing earlier. The purchase price stays where it is.
ElevenLabs alternatives compared
| Tool | How it bills | Entry paid price | Cloning | Offline | If you stop paying |
|---|---|---|---|---|---|
| ElevenLabs | Monthly credits, or API per character | $6 Starter, $22 Creator | Instant from Starter, professional from Creator | No | Unused credits forfeited; clone stays in the account |
| Fish Audio | Monthly credits, or API per byte | $11 Plus | On every plan, free included | No | Plan ends; clone stays in the account |
| Cartesia | Monthly credits, one per character | $5 Pro | Instant from Pro, professional from Startup | No | Plan ends; clone stays in the account |
| OpenAI TTS | API per character | Pay as you go | Approved consent recordings only | No | No access to new renders |
| Kokoro | Free model | $0 | No | Yes, on your hardware | Nothing to stop |
| Chatterbox | Free model | $0 | From a reference clip | Yes, on your hardware | Nothing to stop |
| txt2mp3 | One payment | $29 launch, $99 standing | From a recording, or designed from text | Yes, Apple silicon Mac | Nothing to stop |
Prices were read off each vendor's own page in October 2026. ElevenLabs was running a limited-time Starter offer at the time; the table uses its standing price.
Which alternative fits which job
Real-time voice agents: Cartesia. Speech inside your own code: OpenAI or Fish Audio's API. Cloning on a cloud budget: Fish Audio Plus. A GPU and an afternoon to spare: an open model. One narrator, many minutes a week, on a Mac: a tool you buy once. And for breadth, dubbing and teams, ElevenLabs itself is still the answer.
Before you switch
Count generated minutes, not published ones
Look at last month's usage and divide by what you actually published. That ratio is your retake habit, and every metered price above gets multiplied by it.
Export your takes and keep the reference recording
Credits and cloud clones do not move between services. Your audio files and the original recording do.
Test on your longest script
A voice that impresses on one sentence can drift over ten minutes. Judge every option on the text you actually narrate.
Where ElevenLabs is still the right answer
At low volume, nothing bought outright competes with a free tier or a $6 plan. ElevenLabs is also ahead on the breadth of its voice library, on dubbing, on team seats and on an API built to run at scale, and long-form consistency is still where the cloud leaders lead. If those are what you use, a cheaper meter saves a little and a local app gives up something you need. Stay.
If what you mostly do is narrate, on a Mac, and the bill is the thing you resent, the meter is the problem, and a cheaper meter only shrinks it. The arithmetic per video, at four, twelve and thirty uploads a month, is in the best AI voice for faceless YouTube.
Frequently asked questions
Which ElevenLabs alternatives can also clone a voice?
Fish Audio lists voice cloning on every plan, though only paid plans allow commercial use, and Cartesia offers instant cloning from its $5 Pro plan and professional clones from Startup. OpenAI only offers approved custom voices made from a speaker's consent recording. Among open models, Chatterbox and VoxCPM2 clone from a short reference clip, and txt2mp3 does the same on a Mac. On a cloud service the clone lives in the account; locally it is a file on your disk.
Is there an ElevenLabs alternative that works offline on a Mac?
Yes. txt2mp3 runs the VoxCPM2 speech model on an Apple silicon Mac after a one-time download of about 5 GB, so it renders with the network off and has no per-character meter. Open models such as Chatterbox also run on a Mac through Apple's mps device if you are comfortable setting up Python yourself. Every cloud service on this page needs a connection for each render.
What is the cheapest ElevenLabs alternative that still sounds natural?
By rate, OpenAI's tts-1 at $15 per million characters and Fish Audio's API at $15 per million bytes, about $0.90 for a single take of an hour of English. Fish Audio also clones. Neither is cheapest once retakes and volume grow: every take is billed again, while a one-time purchase stays at the same price however much you render.
Is PlayHT still an ElevenLabs alternative?
No. Meta acquired Play AI in July 2025, and the platform closed on 31 December 2025; reporting on the shutdown says accounts, voice clones and saved audio were deleted with no export tool. Any list that still recommends it is out of date. It is also the clearest example of why a voice held in someone else's account is a risk.
Can I use the Mac's built-in Personal Voice instead?
Not for narration files. Personal Voice is an accessibility feature: it is processed on the device on Apple silicon Macs and speaks through Live Speech and communication apps you allow, but Apple states that those apps can't capture speech from it. It is built for speaking in conversation, not for exporting an MP3 of a script.
Is there an ElevenLabs alternative with a one-time payment?
Yes, though check what the payment buys. Some lifetime deals keep a monthly character allowance or a per-render cap, and some resell a cloud API, so the meter survives the purchase. txt2mp3 runs the model on your own Mac, so there is nothing left to meter: $29 for the launch batch, $99 as the standing price, Apple silicon only.
Where the numbers came from
ElevenLabs pricing
elevenlabs.io/pricingStarter $6 with 30,000 credits and Creator $22 with 121,000 credits, about 121 minutes, $11 for the first month; the extra-minute rate of about $0.18 on Creator; the commercial licence and instant voice cloning from Starter and professional cloning from Creator; rollover of up to two months and at most three times the monthly quota on an active paid plan; that downgrading or cancelling forfeits unused credits; and a limited-time Starter offer running at the time of writing. Fetched 27 October 2026.
ElevenLabs API pricing
elevenlabs.io/pricing/apiText to speech at $0.08 per 1,000 characters on the multilingual models and $0.04 on Flash and Turbo, the same on every tier. Fetched 27 October 2026.
Fish Audio plans
fish.audio/planFree at 8,000 credits, up to 7 minutes, and the FAQ's limit to personal, non-commercial use; Plus at $11 a month for 250,000 credits, up to 200 minutes, commercial use and voice cloning; Pro at $75 for 2,000,000 credits, up to 1,620 minutes; and roughly 600 to 625 credits per minute. Fetched 27 October 2026.
Fish Audio API pricing
docs.fish.audio/developer-guide/models-pricing/pricing-and-rate-limitsThe s2-pro and s2.1-pro models at $15 per million UTF-8 bytes. Fetched 27 October 2026.
Cartesia pricing
cartesia.ai/pricingFree at 20,000 credits; Pro at $5 for 100,000 credits with a commercial licence and instant voice cloning; Startup at $49 for 1.25 million and Scale at $299 for 8 million; professional clones from Startup; one credit per character. Fetched 27 October 2026.
OpenAI API pricing
developers.openai.com/api/docs/pricingtts-1 at $15 per million characters and tts-1-hd at $30. Fetched 27 October 2026.
OpenAI text-to-speech guide
developers.openai.com/api/docs/guides/text-to-speechThe 13 built-in voices, custom voices created from a speaker's consent recording, and the requirement to disclose that the voice is AI-generated. Fetched 27 October 2026.
Kokoro-82M model card
huggingface.co/hexgrad/Kokoro-82M82 million parameters, Apache 2.0, and the v1.0 release's 8 languages and 54 voices. Fetched 27 October 2026.
Chatterbox repository
github.com/resemble-ai/chatterboxMIT licence; Turbo at 350 million parameters, English-only; Multilingual at 500 million and 23 or more languages; cloning from a reference clip; cuda, cpu and mps devices and a Mac example. Fetched 27 October 2026.
VoxCPM2 model card
huggingface.co/openbmb/VoxCPM2Two billion parameters under Apache-2.0, free for commercial use; 30 languages; 48 kHz output; cloning from a short clip and voice design from a description; state-of-the-art or competitive benchmark results; about 8 GB of VRAM. Fetched 27 October 2026.
TechCrunch: Meta acquires voice startup Play AI
techcrunch.com/2025/07/13/meta-acquires-voice-startup-play-aiMeta's acquisition of Play AI, confirmed in July 2025, with the entire team joining Meta.
Texttolab: what happened to Play.ht
texttolab.com/blog/play-ht-shutdown-alternativesThe platform's shutdown on 31 December 2025 and the deletion of accounts, voice clones, saved audio and API keys with no export tools. A secondary report; we found no official notice still online. Fetched 27 October 2026.
Apple Support: Personal Voice
support.apple.com/en-us/104993Personal Voice on Macs with Apple silicon, processed on device, used with Live Speech and allowed communication apps, which can't capture speech from it. Fetched 27 October 2026.
txt2mp3 terms
/termsThat the speech you generate is yours, that we claim no ownership of it, and that you are free to use the audio commercially.
txt2mp3, measured first-hand
The $29, $69 and $99 one-time price ladder; the 120-character trial; the roughly 5 GB setup download; 18.4 seconds to load the model and 5.5 seconds for a short take on our development Mac; 14 preset voices, cloning from a recording and voice design from text; MP3 at 192 kbps or WAV with word-level caption timings; Apple silicon only. The 1,000-characters-a-minute ratio is derived from ElevenLabs' own credits-to-minutes figures, and the hourly costs and break-even months are our arithmetic on the rates above.
