ElevenLabs is the better buy today — and that verdict has an expiry date of 12 October 2026.
Microsoft’s MAI-Voice 2.1, shipped 1 October, lists at $22 per million characters, with MAI-Voice-2.1-Flash at $15. ElevenLabs lists $80 for Eleven v4 and $40 for v4 Turbo. That is the 3-4x gap every roundup is quoting, and it is not the gap you would pay: ElevenLabs is running a 72% launch discount through 12 October that puts v4 at $22 and v4 Turbo at $11 — level with and below Microsoft — while holding the no. 1 slot on Artificial Analysis’ Speech Arena.
Pick ElevenLabs for anything a customer will hear. Pick MAI-Voice-2.1-Flash for high-volume synthesis you can run on preview terms.
The 30-second answer
Pick ElevenLabs if: the audio ships to an audience — narration, audiobooks, ads, dubbing, product voice; you need voice cloning in the EEA or UK; you need documented provenance for published output; you want an SLA; or you are buying before 13 October, when it is also the cheaper option.
Pick MAI-Voice 2.1 / 2.1-Flash if: you are synthesising at API volume where per-character cost dominates, you can tolerate a preview service with no SLA, and the output is internal, ephemeral or regenerable — notifications, IVR prompts, accessibility read-back, generated-on-demand audio.
Pick neither if cost is the only axis: Gemini 3.8 Flash-Lite TTS is cheaper than both until 1 January 2027, and Alibaba’s Qwen-Audio-3.1-TTS is cheaper than everything by an order of magnitude — with no provenance marking, on a China-hosted node.
The three windows this decision lives in
The reason this comparison is unusually time-sensitive is that both vendors moved inside the same fortnight, in opposite directions, and two published dates reshuffle the ranking before the end of the year.
| Window | Cheapest of the two | Why |
|---|---|---|
| 3 – 12 Oct 2026 | ElevenLabs — v4 Turbo at $11/1M | 72% launch discount live on the v4 generation |
| 13 Oct – 31 Dec 2026 | MAI-Voice — Flash at $15/1M vs Turbo at $40 | ElevenLabs reverts to list; Microsoft’s rate unchanged |
| 1 Jan 2027 onward | MAI-Voice, and by more | Google’s TTS rates double, removing the cheap third option; Microsoft’s carries no published expiry |
Two things follow. First, a pilot that finishes before 13 October produces cost numbers you cannot reproduce — build the business case on $80 and $40, not $22 and $11. Second, Microsoft’s voice prices are the only ones in this category right now with no published end date, which is worth something even though an unstated price is not a fixed price. Microsoft already demonstrated the pattern it is capable of: the MAI-Transcribe-2-Streaming rate launched alongside these voices is explicitly introductory and expires 31 December 2026 with no successor published.
Price per hour of generated audio
Per-million-character rates are the wrong unit for a buyer, because nobody budgets in characters. Converted at 900 characters per minute of speech — roughly 150 words a minute at six characters a word including spaces — one hour of finished audio is 54,000 characters.
| Model | Per 1M chars | Per hour of audio | Note |
|---|---|---|---|
| Qwen-Audio-3.1-TTS | ~$3.30 | ~$0.18 | China-hosted node; no provenance marking |
| Gemini 3.8 Flash-Lite TTS | — | $0.54 | doubles to ~$1.08 on 1 Jan 2027 |
| Eleven v4 Turbo (to 12 Oct) | $11 | $0.59 | 72% launch discount |
| MAI-Voice-2.1-Flash | $15 | $0.81 | no published expiry |
| Gemini 3.8 Flash TTS | — | $0.81 | doubles to ~$1.62 on 1 Jan 2027 |
| MAI-Voice-2.1 | $22 | $1.19 | no published expiry |
| Eleven v4 (to 12 Oct) | $22 | $1.19 | 72% launch discount |
| Eleven v4 Turbo / Flash v2.5 / v3 Conversational | $40 | $2.16 | list rate from 13 Oct |
| Eleven v4 / v3 / v2 Multilingual | $80 | $4.32 | list rate from 13 Oct |
Prices verified 3 October 2026 against Microsoft AI’s launch post, the ElevenLabs API price list and Google’s Gemini API pricing page.
Three details in that table are easy to miss. MAI-Voice-2.1 and the discounted Eleven v4 are exactly level at $22 per million characters, which makes the flagship choice purely about quality, terms and compliance for the next nine days. MAI-Voice-2.1-Flash and Gemini 3.8 Flash TTS are exactly level at $0.81 per hour. And MAI-Voice-2.1-Flash is not the cheapest credible Western voice model — Gemini 3.8 Flash-Lite TTS is, by a third, until Google’s published doubling on 1 January 2027 inverts that.
Three vendors, three different meters
The reason published comparisons of this category disagree so wildly is that the three vendors do not bill on the same unit, and converting between them requires an assumption nobody states.
- Microsoft bills per character — $15 and $22 per million.
- ElevenLabs bills per character on the API, and per credit on subscriptions, where website and Studio generation costs 1 credit per character.
- Google bills per audio output token — $9.00 and $6.00 per million tokens for Flash and Flash-Lite TTS, which Google itself also expresses as $0.00225 and $0.0015 per 10 seconds of audio.
Google’s own per-second figure is the only one of the three that needs no assumption, which is why the table above routes everything through hours of audio. Any per-million-character figure quoted for a Gemini TTS model is somebody’s speech-density estimate wearing a price tag, and estimates in circulation differ by more than 2x.
The ElevenLabs credit route deserves its own warning, because it is the one most buyers actually land on. At 1 credit per character, the $990 Business plan’s 6M credits work out to about $165 per million characters — roughly twice the v4 API list rate, and eleven times the MAI-Voice-2.1-Flash rate. API generations are billed against the API price list rather than at the website’s 1-credit-per-character rate, so the route you buy through changes the price more than the vendor you buy from. If you are generating programmatically, confirm you are on API billing before modelling anything.
Where MAI-Voice 2.1 wins
Per-character cost from 13 October. Flash at $15 against v4 Turbo at $40 is a 2.7x saving; MAI-Voice-2.1 at $22 against Eleven v4 at $80 is 3.6x. At 1,000 hours of synthesised audio a month — a mid-size IVR estate or an accessibility read-back feature — that is $810 against $2,160 on the fast pairing, a difference of $16,200 a year.
A rate with no published expiry. Every other cheap option in this category has a documented end date. Microsoft’s two voice models do not. That is an asymmetry rather than a guarantee.
Latency, on the vendors’ own numbers. Microsoft publishes 150 ms end-to-end for MAI-Voice-2.1-Flash. ElevenLabs publishes figures for the v4 generation in a 100-150 ms band. These are three different measurement definitions across two vendors, so treat latency as a tie on paper and measure it on your own traffic if it matters.
Breadth where it counts for voice agents. 23 languages and 26 locales with a single voice carried across all of them — fewer languages than ElevenLabs’ 90+, but consistent voice identity across locales is the harder problem for a multi-region product.
Where ElevenLabs wins
Measured output quality. Eleven v4 took no. 1 on Artificial Analysis’ Speech Arena at launch on 28 September with an Elo of 1319, ahead of Cartesia Sonic 3.6 (1276) and Gemini 3.8 Flash TTS (1267), with reported blind-preference win rates of 65-81% against competing models. Microsoft published no blind-preference ranking for MAI-Voice 2.1. For audio an audience will hear, that gap in evidence is the decisive one.
Terms that exist. ElevenLabs is a generally available commercial product with paid tiers, commercial licensing and enterprise agreements including HIPAA BAAs. Microsoft’s audio models are public preview with no SLA.
Cloning you can lawfully run in Europe. Cloning works in the EEA and UK, Instant Voice Clones need about 10 seconds of audio, Professional Voice Clones are gated behind a consent passphrase spoken in the voice being cloned, and provenance for published output is documented. Compare Google, whose voice replication is blocked in the EEA, UK, Switzerland, India, Illinois and Texas.
The product, not just the model. Voice library, Studio, Dubbing, voice design, sound effects, music at $0.15/minute, and Scribe speech-to-text at $0.22 per hour. MAI-Voice 2.1 is an endpoint. If your workflow is a person producing audio rather than a server generating it, this difference is the whole comparison — which is why ElevenLabs remains the pick on the best AI audio tools list regardless of the rate card.
90+ languages, up from 70 in v3.
The decision table
| Your situation | Pick | Why |
|---|---|---|
| Audiobook, ad, podcast, anything published | ElevenLabs | Measured quality lead; licensing; provenance |
| Cloning a voice in the EU or UK | ElevenLabs | Available and consent-gated; Google is blocked, Microsoft unconfirmed |
| IVR prompts, notifications, accessibility read-back at volume | MAI-Voice-2.1-Flash | Regenerable output absorbs preview risk; 2.7x cheaper from 13 Oct |
| Buying in the next nine days | ElevenLabs | Cheaper and better until 12 Oct |
| Need an SLA | ElevenLabs | Microsoft’s models are preview, no SLA |
| Cheapest provenance-marked rate, any vendor | Gemini 3.8 Flash-Lite TTS | $0.54/hr — until it doubles on 1 Jan 2027 |
| Building a live conversational agent | Neither yet | Price full-duplex alternatives first — different meter, different model |
| Editing spoken audio rather than generating it | Neither | Descript is the tool for that job |
The verdict
For anything that reaches an audience, ElevenLabs is the answer, and the reasoning does not change on 13 October — it only gets more expensive. The v4 generation reclaimed a measured quality lead it had lost to Google and Cartesia through the summer, cloning is lawful and consent-gated in the EEA and UK, provenance is documented, and the terms are real. The price gap that opens on 13 October is the cost of all of that, and for published audio it is worth paying.
For MAI-Voice-2.1-Flash, the case is narrower and genuinely strong inside it: machine-generated speech at volume, where the output is regenerable, no customer is judging the performance, and $16,200 a year is a real number. Take the preview terms seriously — a voice model that shifts character mid-catalogue is audible in a way a changed text completion is not — and confirm regional cloning availability and provenance marking with Microsoft before any of it reaches a published pipeline.
The one thing to avoid is deciding on the 6.7x gap still circulating in roundups. It compares Microsoft’s cheapest model against ElevenLabs’ second-newest at a rate ElevenLabs no longer charges. The real gap is zero this week and about 2.7x next week, and the difference between those two framings is the difference between a sound decision and an expensive one.
What changed
- 3 Oct 2026: Page published. ElevenLabs’ 72% v4 launch discount ($22 / $11 per 1M chars) runs to 12 October; list rates of $80 / $40 resume 13 October.
- 1 Oct 2026: Microsoft shipped MAI-Voice-2.1 at $22 and MAI-Voice-2.1-Flash at $15 per 1M characters, as public preview with no SLA — see the launch analysis.
- 28 Sep 2026: Eleven v4 and v4 Turbo shipped with 90+ languages and took no. 1 on Artificial Analysis’ Speech Arena at 1319 Elo, reclaiming the quality lead from Gemini 3.8 Flash TTS.
MAI-Voice 2.1 vs ElevenLabs — frequently asked questions
Is MAI-Voice actually cheaper than ElevenLabs?
Not until 13 October 2026. Microsoft lists MAI-Voice-2.1 at $22 per million characters and MAI-Voice-2.1-Flash at $15. ElevenLabs' list rates are $80 per million characters for Eleven v4 and $40 for v4 Turbo — but its own API price list shows both discounted 72% until 12 October, to $22 and $11. So right now v4 Turbo at $11 undercuts MAI-Voice-2.1-Flash at $15, and Eleven v4 at $22 is exactly level with MAI-Voice-2.1. On 13 October the discount lapses and Microsoft becomes about 2.7x cheaper on the Flash-versus-Turbo pairing and 3.6x cheaper flagship-to-flagship. The widely quoted '6.7x gap' comes from comparing MAI-Voice-2.1-Flash against a $100 per million figure for the older Eleven v3, which is neither the current model nor the current rate.
What changes on 12 October 2026?
ElevenLabs' launch discount on the v4 generation expires and the API rate card reverts to $80 per million characters for Eleven v4 and $40 for v4 Turbo. Nothing about the models changes; only the price does. If you are running a pilot and comparing invoices, a pilot that ends before 13 October will show you numbers you cannot reproduce afterwards. Build the comparison on the list rates of $80 and $40, treat the discount as a one-off saving on the first few weeks of volume, and if the discounted price is what makes ElevenLabs affordable for your workload then ElevenLabs is not actually affordable for your workload.
Can I put MAI-Voice 2.1 in production?
Microsoft's launch post names no service-level agreement and no general-availability date, and reporting of the Microsoft Foundry listings describes the new audio models as public preview, without an SLA, and not recommended for production. That is the binding constraint on this comparison, and it is more important than the price. A preview model can change voice characteristics, change price or withdraw with no contractual remedy — and for synthesised speech that is a specific problem, because a voice that shifts mid-catalogue is audible to your listeners in a way a changed text completion is not. Pilot on it. Do not re-record a back catalogue on it.
Which one sounds better?
ElevenLabs, on the only public blind-preference benchmark either vendor cites. Eleven v4 took the no. 1 slot on Artificial Analysis' Speech Arena at launch with an Elo of 1319, ahead of Cartesia Sonic 3.6 at 1276 and Gemini 3.8 Flash TTS at 1267, and ElevenLabs reports winning 65-81% of blind head-to-head preference tests against competing models. Microsoft published latency, price, and language coverage for MAI-Voice 2.1 and no blind-preference ranking at all. That is not evidence the model sounds bad; it is an absence of evidence that it sounds good, and for anything a customer will hear, an absence of evidence is a reason to generate your own samples before committing.
Can I clone a voice in the EU or UK with either of these?
ElevenLabs, yes — and this is the clearest practical split in the category. ElevenLabs runs voice cloning in the EEA and UK, gates every Professional Voice Clone behind a spoken consent passphrase recorded in the voice being cloned, and documents provenance for published output. Microsoft's post says MAI-Voice 2.1 clones from a few seconds of reference audio with built-in consent guardrails, but does not set out regional availability or provenance marking, so confirm both with Microsoft before building a publishing pipeline on it. Google is the one to rule out outright here: Gemini's voice replication is unavailable in the EEA, the UK, Switzerland, India, Illinois and Texas.
What is the genuinely cheapest text-to-speech API right now?
Among models with documented provenance and a non-China-hosted endpoint, Gemini 3.8 Flash-Lite TTS — about $0.54 per hour of generated audio, below MAI-Voice-2.1-Flash's $0.81. But Google's rate is documented as introductory and doubles on 1 January 2027, to roughly $1.08 per hour, at which point MAI-Voice-2.1-Flash becomes the cheapest Western rate with no published expiry. Alibaba's Qwen-Audio-3.1-TTS is cheaper than all of them by roughly 4.5x at about $0.18 per hour, and ships no provenance marking on a China-hosted node. The honest answer is that the cheapest option changes twice in the next 90 days, which is an argument for keeping your synthesis layer swappable rather than for picking a winner.