ElevenLabs Alternatives (2026): The Per-Minute Price Ladder, Routed by Job

TL;DR: An ElevenLabs Creator minute costs about $0.18. OpenAI's gpt-4o-mini-tts does a minute for about $0.015, Amazon Polly standard for $0.004, Cartesia for around $0.03 — but every cheap option gives up voice cloning, and nothing on this list matches ElevenLabs dubbing. PlayHT, still recommended by half the roundups for this query, shut down December 31, 2025. Route by job: OpenAI or Polly for bulk narration, Cartesia or Deepgram for realtime agents, Murf for no-code studio work, Kokoro or Chatterbox (never XTTS v2) for commercial self-hosting.

Half the "best ElevenLabs alternatives" roundups in this SERP still recommend PlayHT. PlayHT no longer exists. Meta acquired the team in July 2025, the service was permanently shut down on December 31, 2025, and accounts, voice clones, and API access were deleted with no export path (third-party reports: infrabase.ai, theplanettools.ai, notevibes). We confirmed the death ourselves in July 2026 the simplest way possible: play.ht and www.play.ht no longer resolve — our fetches returned DNS failures. Any list that ranks it is telling you when it was last updated.

So this page starts from live pages instead. Here is the baseline you are pricing against, verified live on elevenlabs.io/pricing in July 2026 (vendor prices): Free at $0 for 10,000 credits, Starter $6 for 30,000, Creator $22 for 121,000, Pro $99 for 600,000, Scale $299 for 1.8 million, Business $990 for 6 million. One credit buys one character of standard text-to-speech, and roughly 1,000 characters make one minute of finished speech — the industry rule of thumb ElevenLabs ' own credit system implies, and the assumption behind every per-minute number on this page. That prices a Creator minute at about $0.18 and a Pro minute at about $0.165 (our math from those vendor prices). The "low-latency TTS as low as 5c/minute" line on the same page is real, but it is a Business-tier claim, and Business starts at $990 a month. Two more line items from the same page: dubbing burns 2,000 to 10,000 credits per minute, and annual billing works out to paying for ten months.

Why people leave is consistent across user reports (Reddit threads surfaced via the cloudtalk.io and brilo.ai roundups — user reports, not our tests): regenerations bill the full text, so changing one word means paying for the whole paragraph re-render; credits do not roll over, and cost visibility is poor; non-English voices and accents come out inconsistent; and support is described as a chatbot wall, with some paying customers reporting automated account bans. None of that means the audio is bad. It means the bill and the platform are the problem, and both are fixable by switching — if you know which job you are hiring a voice engine for.

Disclosure: we have no affiliate or business tie to ElevenLabs or any vendor on this page as of publication; if that changes, this line will say so. We have not run our hands-on audio-quality suite on these engines yet; prices and features come from the live vendor pages cited below, third-party trackers where a vendor page blocked our fetcher (each such number is labeled), and labeled user reports. We track LLM API prices live with the same methodology.

Every alternative at a glance

Assumption for every per-minute figure: ~1,000 characters ≈ 1 minute of finished speech. Back-computed cells are our math from vendor prices.

The per-minute price ladder

This is the table the stale roundups don't print. Same assumption as above: ~1,000 characters ≈ 1 minute of finished speech; every per-minute cell is our math from the cited vendor prices.

Read the ladder honestly: the 10x price gap between the API providers and ElevenLabs is real, but so is the pattern in the cloning column. Everything under $0.04 a minute either has no cloning at all or (Cartesia) a much thinner voice library. You are not finding " ElevenLabs but 10x cheaper." You are deciding whether your job needs what the premium buys. If you like this format, the same cost-ladder logic for text models is in our cheapest LLM API breakdown .

OpenAI TTS API: the budget default for narration

The cheapest mainstream API per minute, at roughly a tenth of an ElevenLabs Creator effective minute. Rates: gpt-4o-mini-tts at $0.60 per million text input tokens plus $12 per million audio output tokens, which works out near $0.015 per minute; tts-1 at $15 per million characters (~$0.015/min); tts-1-hd at $30 per million (~$0.03/min). One sourcing caveat we owe you: these are vendor-published rates, but OpenAI's pricing pages returned 403 to our fetcher in July 2026, so we cross-checked against three third-party trackers (costgoat, texttolab, tokenmix) and the OpenAI community forum — all consistent. Label them vendor rate, third-party-confirmed, not verified-live.

What you get: pay-as-you-go with zero subscription, and "instructable" styling on gpt-4o-mini-tts — you can tell it to speak like a calm narrator or a hyped sportscaster. If your app already calls OpenAI, this is the obvious first test.

The catch: no voice cloning of any kind — thirteen preset voices, full stop. No dubbing product. Voices are good-not-great for character work. And the audio-token billing has a gotcha users flag on the OpenAI developer forum (user report): slower, drawn-out speech generates more audio tokens, so the same word count can cost more depending on delivery. Budget with margin.

Cartesia Sonic: the realtime and voice-agent pick

Cartesia is the only credible cloning-plus-API rival to ElevenLabs on this list, and it is built for the job ElevenLabs prices vaguely: low-latency streaming agents. Verified live on cartesia.ai/pricing in July 2026 (vendor prices): Free at $0 for 20,000 credits; Pro at $5/month for 100,000 credits; Startup at $49 for 1.25 million; Scale at $299 for 8 million. Voice-agent calls are a flat $0.06 per minute, and a Cartesia phone number adds $0.014 per minute of telephony — line items ElevenLabs does not itemize at all.

One transparency note: Cartesia's pricing page never defines what one credit equals. The per-minute figures we print (~$0.038 on Pro, ~$0.029 on Startup, ~$0.028 on Scale) are back-computed from the page's own minutes claims — that math is ours, not the vendor's. Instant voice cloning ships at the $5 tier — a near-match for ElevenLabs , which includes instant cloning from its $6 Starter plan; professional-grade cloning actually favors ElevenLabs , at $22 (Creator) versus Cartesia's $49 Startup tier.

The catch: a smaller voice library and a younger ecosystem than ElevenLabs , and there is no dubbing or audiobook production suite — this is an engine, not a studio. If the job is a phone agent, start with our AI receptionist guide , where the latency and telephony math gets applied end to end.

Deepgram Aura-2: TTS and STT on one bill

Deepgram's angle is that a voice agent is only half text-to-speech — it also has to listen. Aura-2 runs $0.030 per 1,000 characters pay-as-you-go, $0.027 on the Growth plan (verified live on deepgram.com/pricing, July 2026, vendor price); the older Aura-1 is $0.015 per 1,000. That is simple per-character pricing: no credit system, no subscription, and speech-to-text from the same vendor on the same bill, which is exactly what a phone-agent stack wants.

The catch: no voice cloning, a small voice selection, and enterprise-utilitarian voice quality — Aura-2 is tuned for IVR and agent clarity, not audiobook warmth. Nobody narrates a novel with it. But if the job is "answer the phone and be understood," it is arguably the more honest tool.

Amazon Polly: cheapest at scale, oldest voices

Verified live on aws.amazon.com/polly/pricing in July 2026 (vendor prices): standard voices at $4 per million characters (~$0.004/min — roughly 45x cheaper than an ElevenLabs Creator minute), neural at $16 per million (~$0.016/min), generative at $30 per million (~$0.030/min), and a long-form engine tuned for narration at $100 per million (~$0.10/min). The free tier is genuinely large: 5 million standard characters a month indefinitely, plus 1 million neural, 500,000 long-form, and 100,000 generative characters a month for the first twelve months.

Who it fits: teams already on AWS who want IAM, regions, and SLAs around their TTS, and anyone generating speech at a volume where $0.004 versus $0.18 a minute is the whole business case.

The catch: no voice cloning — Brand Voice is an enterprise engagement with AWS, not a product you can buy; standard and neural voices sound dated next to 2026-generation models; and the generative voices that do compete on quality erase most of the price advantage over OpenAI. Non-developers will also find the AWS console a hostile place to make an MP3.

Murf AI: the studio for people who will never call an API

Murf fills the slot none of the APIs above touch: a browser studio with timeline editing, slide sync, and team seats, aimed at marketing, e-learning, and corporate video. Murf even runs an official PlayHT-refugee migration guide — a clear signal about which user base it is absorbing.

Pricing carries a mandatory label here: Murf's own pricing page is JS-rendered and returned nothing to our fetcher, so these tiers are third-party reports (fish.audio, checkthat.ai, G2 pricing pages, searched July 2026), not live-verified. Creator at roughly $29/month ($19/month billed annually) for about two hours of generation a month — an effective $0.16 to $0.24 per minute — and Business at roughly $99/month ($66 annual). Third parties also flag an oddity worth confirming with sales: monthly-billed Business reportedly includes more hours per year than annual-billed.

The catch: you are paying ElevenLabs -class prices or worse per minute — the money buys the studio workflow, not cheaper audio. Hours do not roll over, voice cloning sits in higher tiers or enterprise, and there is no serious API or realtime story.

Self-hosting: Kokoro, Chatterbox, Piper — and the XTTS v2 licensing trap

Open-source TTS is the only route to unlimited generation, total privacy — voice data never leaves your infrastructure, which reads differently after this spring's biometric-privacy suits (below) — and voice cloning with no platform terms of service. Third-party reports put self-host costs well under $0.01 per minute at decent GPU utilization, but treat any such number as unverified: it depends entirely on your hardware and load.

Now the trap, because most roundups miss it and it has teeth. XTTS v2 is still widely called the best local multilingual voice-cloning engine in 2026 — and its model weights are licensed under the Coqui Public Model License, which is non-commercial, full stop (license terms verified via Hugging Face and the maintained GitHub fork, July 2026). Coqui Inc. shut down in January 2024, so there is no company left to sell you a commercial license. Only the surrounding TTS code is MPL-2.0. Shipping XTTS v2 inside a product is not a gray area; it is unlicensable. The commercial-safe picks are Kokoro (Apache-2.0), Chatterbox (MIT), and Piper — with one caveat on Piper: its MIT license covers the original rhasspy/piper engine, while the actively maintained successor fork (piper1-gpl) is GPL-3.0, and individual voice checkpoints carry their own dataset terms, so check the repo and voices you actually ship.

The rest of the catch is ordinary self-host reality: you run the GPUs, there is no support, quality sits below the ElevenLabs and Cartesia flagships, and cloning-abuse liability is entirely yours. Whether that trade makes sense is the exact question our build vs buy guide for AI agents is built to answer.

The consent backdrop every voice buyer should know

This spring moved voice cloning from a product question toward a legal one. Between May 11 and 13, 2026, nine class actions were filed in Chicago federal court under Illinois' Biometric Information Privacy Act against ElevenLabs , Amazon, Apple, Google, Meta, Microsoft, Adobe, NVIDIA, and Samsung, alleging the companies trained voice models on journalists' and narrators' recordings without consent; named plaintiffs include Robin Amer, Lindsay Dorcus, and Victoria Nassif (as reported by Sifted and industry trackers; these are allegations, not findings). Consumer Reports found in March 2025 that most voice-cloning products it examined lacked meaningful safeguards against fraud. For its part, ElevenLabs gates professional cloning behind a spoken-CAPTCHA verification and bans impersonation and election misuse — policies tightened after the January 2024 Biden robocall incident. None of this tells you which engine to buy. It does tell you to keep consent records for any voice you clone, on any platform, and it is one more argument the self-host section's privacy case borrows.

Pick by job

• Audiobooks and long-form narration on a budget: OpenAI gpt-4o-mini-tts (~$0.015/min) or Polly's long-form/generative engines. If the quality ceiling is the product — a flagship audiobook, a branded voice — staying on ElevenLabs is a defensible answer; that is what the $0.18 buys.
• Realtime phone and voice agents: Cartesia (flat $0.06/min agent pricing, instant cloning at $5) or Deepgram (TTS + STT on one bill). Work through the full stack in our AI receptionist guide before committing.
• Dubbing: nobody on this list matches ElevenLabs ' dubbing studio. Budget for it honestly — 2,000 to 10,000 credits per minute — or don't take the job. This is the clearest "stay" verdict on the page.
• Self-host and privacy: Kokoro or Chatterbox for anything commercial; XTTS v2 only for personal or research use. Decision framework: build vs buy .
• Non-developer studio work: Murf. You will not save money over ElevenLabs ; you will save yourself an API.
• Stranded PlayHT customer: your clones are gone — no export ever existed. Nearest matches: Cartesia for API and agent work, Murf (which publishes an official migration guide) for studio work.

Where the alternatives fall short

The honest close: nothing on this page replaces ElevenLabs one-for-one. No alternative matches its combination of clone quality, voice-library breadth, and a real dubbing product in one platform. Every cheap option on the ladder gives up cloning entirely — OpenAI, Deepgram, and Polly all ship preset voices only. Cartesia is the one credible clone-plus-API rival, and it is younger, with a thinner library and a pricing page that won't say what a credit is. Murf costs ElevenLabs money for non- ElevenLabs audio. Self-hosting trades the subscription for GPUs, ops work, and liability. And we repeat the standing caveat: we have not run our hands-on audio-quality suite on these engines yet, so every quality statement above is labeled to its source — vendor claim, third-party report, or user report — rather than our own listening test. If your complaint with ElevenLabs is the bill, the ladder gives you four ways down. If your complaint is quality, wait for our blind test before you move.

All guides in this topic

• Live LLM API pricing tracker — the same verify-it-live methodology, applied to text models daily
• The cheapest LLM API in 2026 — the text-model cost ladder that inspired this page's format
• AI receptionist: the full stack — where the realtime TTS picks above actually get deployed
• Build vs buy for AI agents — the self-host decision, generalized
• OpenRouter review — one API surface across many model vendors, if you'd rather not marry any of them

Frequently asked questions

Частые вопросы