ElevenLabs has spent the last two years becoming the default voice of the AI era — the narration behind millions of YouTube videos, audiobooks and phone agents. On 11 November 2026, at its first New York Summit, the company made its biggest move yet: ElevenLabs Scribe v3, a new speech-to-text model, arrived as the headline of a 13-item package that also includes a new text-to-speech model, a speech-to-speech model, a healthcare-tuned transcription model, an orchestration layer called the Interaction Engine, and a workplace app.
The package is aimed squarely at businesses rather than hobbyists, and about half of it is preview or early access rather than finished product — so this guide separates what you can use today from what is still a roadmap.
ElevenLabs Scribe v3 is the part most readers will actually touch first. It transcribes recorded audio in 102 languages, and ElevenLabs says it makes 27% fewer word errors than Scribe v2 on its own evaluation. Its real-time sibling, Scribe v3 Live, adds speaker labels as people talk. The price has not changed: $0.20 per hour of audio in the API, the same as v2. That combination — better accuracy at the same price — is why this launch matters even if you never buy an enterprise voice agent.
So what exactly does ElevenLabs Scribe v3 do better, what do the other twelve announcements add, how much does the whole stack cost, and should you switch your transcription or voice-agent work to it? This guide breaks down every announcement, the full pricing, where each piece is available today, and the honest caveats ElevenLabs’ own materials quietly include. Where a number comes from the company rather than an independent test, we say so.

Table of Contents
- What Is ElevenLabs Scribe v3?
- ElevenLabs Scribe v3 at a Glance
- ElevenLabs Scribe v3 vs Scribe v2: The Accuracy Numbers
- Scribe v3 Live: Real-Time Transcription With Speaker Labels
- The Other Two Models: Eleven v4 and Eleven Realtime
- Scribe v3 Medical: Transcription Tuned for Healthcare
- The Interaction Engine: The Centrepiece of the Summit
- ElevenAgents Upgrades: Memory, Routines and Architect
- ElevenAI: The New Workplace App
- Custom Models, the Partner Program and the Nvidia Deal
- ElevenLabs Scribe v3 Pricing: The Full Rate Card
- What Is Available Now — and What Is Still Coming
- How to Try ElevenLabs Scribe v3 Today
- Who Should Use ElevenLabs Scribe v3 — and Who Should Wait
- ElevenLabs Scribe v3 vs the Competition
- ElevenLabs Scribe v3 Limitations and Honest Caveats
- ElevenLabs Scribe v3 FAQs
What Is ElevenLabs Scribe v3?
ElevenLabs Scribe v3 is the company’s third-generation speech-to-text model, launched on 10 November 2026, a day before the New York Summit where it was presented as part of the 13-item package. It turns recorded audio — meetings, interviews, lectures, podcasts, voice notes, customer calls — into text, and it does so in 102 languages.
If you have used Scribe v2, the upgrade is a drop-in change: existing integrations switch by changing the model ID to scribe_v3 in the API, and both models are now the default in the ElevenLabs app.
The headline claim is accuracy. ElevenLabs says ElevenLabs Scribe v3 makes 27% fewer word errors than Scribe v2 across 102 languages, measured on an internal FLEURS evaluation set. In the company’s numbers, the word error rate falls from 16.94% to 12.38%, and it improves in every language tested. That is a company-run evaluation, not an independent benchmark, so treat the exact percentage as a direction rather than a guarantee — but a 27% error reduction, if it holds on your own audio, is the difference between a transcript you can search and quote and one you have to re-listen to.
What makes ElevenLabs Scribe v3 interesting beyond the headline number is the set of features around the core transcription. It supports speaker diarisation, so a meeting transcript can show who said what. It handles mid-sentence language switching, transcribing each stretch in its own script rather than forcing the whole file into one language. And it offers a keyterm library — a custom vocabulary of product names, technical terms and proper nouns that the model consults when it hears a likely match — which matters enormously for anyone transcribing industry-specific speech, where a generic model turns brand names into gibberish.
Under the hood, ElevenLabs Scribe v3 is also a compliance story. The models run in the US by default, with isolated EU, India and Singapore environments for enterprise customers. A Zero Retention Mode deletes audio and text when a request completes. The service is covered by SOC 2, ISO 27001 and PCI DSS Level 1 programmes, is HIPAA-eligible with a business associate agreement, and is available for private deployment. For teams in regulated industries, that checklist is often the real reason to switch — accuracy gets you in the door, compliance gets you past legal.
ElevenLabs Scribe v3 at a Glance
Put together, ElevenLabs Scribe v3 is three things at once: a more accurate transcription model, a real-time listening layer for voice agents, and a compliance-checked service that regulated teams can actually buy. The recorded model handles the archive — meetings, interviews, podcasts. The Live model handles the moment — calls, captions and agents. And the keyterm library handles the vocabulary that generic models always get wrong. None of those three is new on its own; what is new is getting all three, at the old price, from one vendor, on the same day.
ElevenLabs Scribe v3 vs Scribe v2: The Accuracy Numbers
The comparison that matters most to existing users is simple: what changes when you move from Scribe v2 to ElevenLabs Scribe v3? The answer, according to ElevenLabs, is fewer errors everywhere, at the same price, with no code changes beyond the model ID.
| Capability | Scribe v2 | ElevenLabs Scribe v3 |
|---|---|---|
| Languages | 99 | 102 |
| Word error rate (ElevenLabs’ FLEURS set) | 16.94% | 12.38% |
| Error reduction vs predecessor | — | 27% fewer word errors |
| API price per hour of audio | $0.20 | $0.20 |
| Speaker diarisation | Available | Improved, with real-time labels in Live |
| Mid-sentence language switching | Limited | Multilingual mode in Scribe v3 Live |
| Custom vocabulary | Keyterm prompting | Keyterm library (add-on) |
| Default in the ElevenLabs app | Yes, until now | Yes — v3 is the new default |
A few things are worth noting about these numbers. First, the 27% figure is ElevenLabs’ own measurement on its own test set — the company has not published an independent third-party evaluation of ElevenLabs Scribe v3 yet, and no rival has had time to run one.
Second, the improvement is reported across all 102 languages, which is unusual: most transcription upgrades improve the big languages and quietly regress on smaller ones. If that holds up in practice, it is good news for anyone transcribing speech in languages beyond English, Spanish and Mandarin. Third, the price freeze matters. Transcription is a commodity market where a tenth of a cent per hour is a competitive weapon, and ElevenLabs chose to hold the line at $0.20 per hour rather than charge for the accuracy gain.
The practical test is cheap to run. Because ElevenLabs Scribe v3 costs the same as v2, you can run the same hour of your own audio through both models and compare the transcripts side by side. Pay particular attention to the words that cost you money when they are wrong: names, numbers, product codes and drug names. A model that is 27% better on average but still mangles your company’s product name is not better for you — which is exactly the problem the keyterm library is meant to solve.

Scribe v3 Live: Real-Time Transcription With Speaker Labels
ElevenLabs Scribe v3 handles recorded audio. Its sibling, Scribe v3 Live, handles audio as it arrives — phone calls, live captions, dictation, and the listening half of a voice agent. It launched on the same day, 10 November 2026, and it is the more technically interesting of the two.
The headline for Scribe v3 Live is that it labels speakers in real time, for the first time in the Scribe line. On a live call, the transcript can show who is speaking as they speak, rather than sorting that out afterwards. ElevenLabs says Scribe v3 Live makes 43% fewer errors than Scribe v2 Realtime, a bigger jump than the recorded model’s 27%, which makes sense: real-time transcription is the harder problem, because the model cannot look ahead to the end of a sentence before committing to a word.
Two other features are aimed at real conversations rather than clean recordings. Turn prediction estimates when a speaker has finished, so a voice agent knows when to reply instead of talking over the caller or waiting too long. Multilingual mode transcribes mid-sentence language switches, with each language rendered in its own script — useful for the many real conversations that mix English with Spanish, Hindi, Arabic or another language in a single sentence.
The official product page quotes a median latency of 150 milliseconds to a committed transcript for Scribe v3 Live, which is fast enough for live captions and for agents that need to react while the caller is still talking.
There are two catches to know before you build on it. First, the official SDKs do not yet support turn prediction, multilingual mode or the keyterm library — those features need a direct WebSocket connection for now, which is more engineering work. Second, the keyterm features are billed as add-ons on top of the base $0.22 per hour, so a heavily customised deployment costs more than the headline rate. Neither is a deal-breaker, but both are the kind of detail that turns a smooth demo into a delayed launch if you discover them late.
Scribe v3 Live is available in ElevenAgents, ElevenLabs’ voice-agent platform, and in the ElevenAPI under the model ID scribe_v3_live. One early enterprise tester, the energy company Eneco, ran Scribe v3 Live on real customer calls before launch — a sign that ElevenLabs is pitching the live model first at contact centres and support teams, where a transcript that is wrong in real time is worse than no transcript at all.
The Other Two Models: Eleven v4 and Eleven Realtime
ElevenLabs Scribe v3 listens. The other two models in the Summit package speak — and they are what turn a transcription upgrade into a full voice stack.
Eleven v4: expressive text-to-speech
Eleven v4 is the text-to-speech model that reads a script the way a person would: speeding up, slowing down and shifting tone with the words. It launched on 28 September 2026 on a completely new architecture, and it is the model behind the voices most readers will actually hear — audiobook narration, YouTube voiceovers, video dubbing, and the spoken half of a voice agent. It reads the tone, pacing and context of a script rather than reading each sentence in isolation, and it takes plain-English stage directions written inline, such as [whispers] or [excited], to direct the performance.
The numbers behind it are strong. Eleven v4 ranked first on Artificial Analysis’ Provider Voice Arena, and ElevenLabs reports that listeners preferred it in roughly 75% of blind head-to-head tests against rivals including Cartesia Sonic 3.6, Inworld TTS-2 and Google’s Gemini 3.8 Flash TTS. Its low-latency sibling, Eleven v4 Turbo, answers with a median inference latency of around 100 milliseconds and is built for real-time agents.
API pricing is $0.08 per 1,000 characters for the full model and $0.04 per 1,000 for Turbo. We covered the model in depth when it launched — see our ElevenLabs v4 review for the full benchmark breakdown, the audio-tag system and the pricing tables.
Eleven Realtime: speech-to-speech voice agents
Eleven Realtime is a newer class of model: speech in, speech out, with no text step in the middle. It launched on 20 October 2026 in two versions, v1 Mini and v1 Max, and it is built for voice agents that need to respond fast and naturally — the kind of agent that answers a phone call, not the kind that narrates a video. It runs in ElevenAgents and through an OpenAI-compatible Realtime API, priced at $0.08 per minute of connected audio plus the cost of the language model behind it.
The distinction matters for buyers. Eleven v4 is for produced speech: you write a script, the model performs it. Eleven Realtime is for conversational speech: the caller speaks, the model listens, thinks and replies in one continuous loop. ElevenLabs Scribe v3 sits underneath both, because every voice agent needs to hear accurately before it can answer well — and Scribe v3 Live is the listening layer inside ElevenAgents. For the head-to-head with Mistral’s competing voice stack, see our Voxtral 3 vs Eleven Realtime comparison.
Scribe v3 Medical: Transcription Tuned for Healthcare
Alongside the general model, ElevenLabs launched Scribe v3 Medical, a transcription model tuned for healthcare terminology. The company’s own example is telling: it distinguishes cycloserine, a tuberculosis drug, from cyclosporine, an immunosuppressant — two words that sound nearly identical and mean dangerously different things. In a clinical note, that is not a typo; it is a patient-safety issue.
Scribe v3 Medical is available now, with a variant tuned for medical conversations inside ElevenAgents. ElevenLabs says it plans to build more industry-specific models, with financial services named as the next sector. The strategy is clear: transcription accuracy is becoming a commodity, so the company is competing on the last mile — the vocabulary of a specific job — where a generic model, however good, will always lose to a tuned one.
For US buyers, the compliance picture matters as much as the accuracy. Scribe models are HIPAA-eligible with a business associate agreement, support Zero Retention Mode, and can run in isolated enterprise environments. That combination is what lets a hospital or clinic send dictation to a cloud transcription service at all.
But eligibility is not a guarantee: any healthcare team still needs its own BAA in place, its own data-handling review, and its own testing on real clinical speech — including accents, background noise and the abbreviations doctors actually use — before trusting a transcript in a medical record. A model that gets drug names right and still mishears a dosage is not safe to deploy.
The Interaction Engine: The Centrepiece of the Summit
If ElevenLabs Scribe v3 is the headline, the Interaction Engine is the argument. It is a new orchestration layer that sits across a whole voice conversation and decides which model should handle each moment. A routine question — checking an order status, confirming an appointment — can go to a cheaper model. A complex or emotional moment goes to a stronger one. The engine manages turn-taking and interruptions, so the agent knows when to speak and when to wait, and it hands the conversation between models without the caller noticing.
The idea is simple and the economics behind it are not trivial. Voice is where AI agents meet customers first — on the phone, in support chats, in meetings — and it is also where their costs are most visible, because every second of conversation burns tokens twice: once to hear, once to think. A voice agent that runs a frontier model on every word is accurate and unaffordable. One that runs a cheap model on every word is affordable and wrong at the worst moments.
The Interaction Engine is ElevenLabs’ bet that the winner in voice will not be the company with the single best model, but the one that orchestrates many models well and remembers who it is talking to.
ElevenLabs’ co-founder and chief executive, Mati Staniszewski, put it this way at the Summit: the Interaction Engine “orchestrates all this under the hood, so you can be confident you’re delivering the right outcome for the customer, without wasting money on tokens.” What the company has not disclosed is any independent benchmark showing how much businesses could actually save, or whether the system improves customer resolution rates compared with existing deployments. That is the honest gap in the announcement — and the reason to treat the Interaction Engine as a promising preview rather than a proven saving.
That preview status matters. The Interaction Engine launched in preview at the Summit, not as a generally available product, and it is aimed at businesses first. Most of the new items in the package are previews, early access or phased rollouts rather than finished products you can use today. So the immediate effect is a roadmap for enterprise buyers, not a new catalogue — and the right response is to test it on your own real calls, measure what it actually saves, and judge the claims against your own results before you pay or switch.

ElevenAgents Upgrades: Memory, Routines and Architect
ElevenAgents is ElevenLabs’ platform for building voice and chat agents, and the Summit brought it three upgrades that change what an agent can do across more than one conversation.
Memory gives agents the full context of every returning customer, recalling what they have said before — from a new shirt size to a pending house move — so nobody has to repeat themselves. This is the feature customers actually feel: the difference between a support line that asks for your account number every time and one that picks up where the last call left off.
It is also the feature that raises the sharpest privacy questions, because memory is only useful if the agent stores what you said. ElevenLabs’ answer is the same compliance stack as the rest of the platform — Zero Retention Mode, isolated environments, SOC 2 and ISO 27001 — but any team turning Memory on should decide explicitly what gets remembered, for how long, and who can delete it.
Routines add multi-step automation: an agent that does not just answer a question but carries out the follow-up work — filing the ticket, sending the confirmation, updating the record — without a human in the loop for each step. Architect is the third piece, a tool for designing and structuring those agents. Together with Campaign Manager, an agent-orchestration tool that shipped on 9 November, the day before Scribe v3 arrived, they turn ElevenAgents from a place to build a single voice bot into a place to run a fleet of them.
The platform also gained a new reasoning option in the same week: on 11 November, ElevenLabs and Nvidia announced that Nvidia’s Nemotron 3.5 Super VL model is available in ElevenAgents as the reasoning model behind a voice agent. That is part of a wider expanded Nvidia partnership, which also includes 15 megawatts of new Vera Rubin compute capacity for ElevenLabs — infrastructure that matters because voice agents are compute-hungry, and capacity is what keeps them answering during peak hours.
ElevenAI: The New Workplace App
The most consumer-facing piece of the package is ElevenAI, a new app that acts as an AI teammate for enterprise teams, starting with meetings. It takes notes, writes summaries and pulls out action items — the same job as the meeting assistants that have become a crowded category in 2026 — but with ElevenLabs Scribe v3 doing the listening underneath, which is a genuine advantage if the accuracy claims hold. A meeting app lives or dies on whether its transcript can be trusted, and a 27% error reduction is the kind of improvement a user notices within a week.
ElevenAI opened in free early access at the Summit, which means it is usable now but not finished: expect rough edges, missing integrations and a feature list that changes quickly. For teams already paying for a meeting assistant, the honest question is whether ElevenAI’s transcription is enough better to justify switching tools mid-year. For teams still taking notes by hand, the free early access is a low-risk way to find out. If you want a broader view of the category first, our comparison of the best AI meeting assistants in 2026, with real prices, puts ElevenAI’s entry in context.
Custom Models, the Partner Program and the Nvidia Deal
The most enterprise-flavoured announcements are the ones with the least detail. ElevenLabs will now tune custom models on each enterprise customer’s own live data — a model that knows your products, your terminology and your customers’ phrasing, rather than a general model that guesses. But the work has not started yet: it begins in the coming months, with selected enterprises first, and no pricing has been disclosed. Do not present custom-model tuning as something any customer can sign up for today, because it is not.
The go-to-market side of the package is more concrete. A new Partner Program counts more than 550 partners in 44 countries, backed by a Marketplace where partners can list what they build. ElevenLabs also announced an alliance with McKinsey to help organisations build and scale voice agents — the kind of partnership that matters less for the technology than for the signal it sends to large buyers, who want a consultancy standing behind a vendor before they sign a multi-year deal.
The Nvidia partnership deserves a line of its own. Beyond the Nemotron model in ElevenAgents and the 15 megawatts of Vera Rubin capacity, it positions ElevenLabs as an infrastructure-scale voice company rather than a model shop. Voice agents need GPUs the way call centres need phone lines, and a supplier that can guarantee capacity wins deals that a supplier renting spare compute cannot.
For readers, the practical takeaway is narrower: ElevenLabs is investing in the capacity to keep its agents running at scale, which is a good sign for reliability, and a reminder that the voice-agent market is consolidating around a few large platforms.
ElevenLabs Scribe v3 Pricing: The Full Rate Card
The best news in the whole package is the price list — or rather, what did not change. ElevenLabs Scribe v3 costs the same as Scribe v2 in the API: $0.20 per hour of audio. Scribe v3 Live costs $0.22 per hour, the same as Scribe v2 Realtime. In a market where every model upgrade usually arrives with a price rise, holding the rate flat while claiming a 27% accuracy improvement is a deliberate competitive move, and it is the strongest reason for existing users to switch: the upgrade costs nothing.
| Product | Price | Notes |
|---|---|---|
| ElevenLabs Scribe v3 (API) | $0.20 per hour of audio | Unchanged from Scribe v2 |
| Scribe v3 Live (API) | $0.22 per hour of audio | Unchanged from Scribe v2 Realtime |
| Scribe v3 in the web app | 330 credits per minute | Draws on your plan’s monthly credit pool |
| Keyterm add-ons | Extra per hour | Keyterm prompting and library billed separately |
| Eleven v4 (API) | $0.08 per 1,000 characters | Expressive text-to-speech |
| Eleven v4 Turbo (API) | $0.04 per 1,000 characters | Low-latency, ~100ms median inference |
| Eleven Realtime (API) | $0.08 per minute | Plus LLM tokens; speech-to-speech agents |
| ElevenAgents platform | $0.08 per call minute | Plus LLM and phone carrier costs |
Subscription plans all draw on one monthly credit pool, and speech costs one credit per character in the web app.
The tiers are: Free at $0 per month with 10,000 credits (about 10 minutes of audio); Starter at $6 per month with 30,000 credits; Creator at $22 per month with 121,000 credits; Pro at $99 per month with 600,000 credits; Scale at $299 per month with 1,800,000 credits; and Business at $990 per month with 6,000,000 credits. Annual billing is ten times the monthly price — two months free. Prices exclude tax, and the Starter plan’s first month is $1 on the current promotion.
For voice agents, the plan sets your included minutes and how many calls can run at once. The Free plan includes 15 minutes a month across 4 simultaneous calls; Starter includes 75 minutes; Creator 275 minutes; Pro 1,238 minutes; Scale 3,738 minutes; Business 12,375 minutes. The language model behind the agent is billed at the provider’s published rate with no markup, and phone carrier charges sit on top.
A quick sense-check on cost: one hour of ElevenLabs Scribe v3 transcription costs $0.20 — about a third of the $0.36 per hour that OpenAI’s Whisper API charges at $0.006 per minute. For a team transcribing a hundred hours of calls a month, that is $20 against $36, before any accuracy difference is counted.

What Is Available Now — and What Is Still Coming
About half the Summit package is shipping and about half is a roadmap. Here is the honest split, because confusing the two is the easiest way to be disappointed.
| Item | Status |
|---|---|
| ElevenLabs Scribe v3 | Available now — default in the app and in the API |
| Scribe v3 Live | Available now — in ElevenAgents and the API |
| Scribe v3 Medical | Available now, including a variant in ElevenAgents |
| Eleven v4 and v4 Turbo | Available now — shipped in the month before the Summit |
| Eleven Realtime | Available now — shipped in the month before the Summit |
| ElevenAI workplace app | Free early access — usable, not finished |
| ElevenAgents Memory and Routines | Announced — rolling out to customers |
| Campaign Manager | Available — shipped 9 November 2026 |
| Interaction Engine | Preview — enterprise-first, no GA date |
| Custom models on customer data | Not started — work begins in coming months, selected enterprises first |
| Financial-services model | Planned only — no date |
| Partner Program and Marketplace | Open — 550+ partners in 44 countries |
The timeline tells its own story. Campaign Manager shipped on 9 November. ElevenLabs Scribe v3 and Scribe v3 Live arrived on 10 November. The Summit itself was 11 November. Three days of announcements in a row, with the models landing before the stage presentation — which is why the reporting around the Summit reads like a product launch rather than a keynote teaser. The models are real and shipping; the orchestration and customisation layers are the work in progress.
How to Try ElevenLabs Scribe v3 Today
Getting started takes minutes, and the free tier is enough for a real evaluation. Here is the practical path:
- Create a free account at the official ElevenLabs site — no credit card is required for the 10,000-credit free tier, which covers about 10 minutes of audio.
- Upload a recording you already have — a meeting, an interview, a voice note — and transcribe it with ElevenLabs Scribe v3. It is the default model in the app now, so you do not need to change a setting.
- Run the same file through your current tool and compare the two transcripts word for word. Count the errors that matter to you: names, numbers, product terms and drug names, not filler words.
- Test the keyterm library if your audio has specialist vocabulary. Add your product names and technical terms, then re-run the file and see whether the errors disappear.
- Try Scribe v3 Live if you build voice agents or need live captions. It is in ElevenAgents and in the API as scribe_v3_live, and the real-time speaker labels are the feature to test first.
- Switch your integration only after your own test: change the model ID to scribe_v3 or scribe_v3_live in the API. Existing Scribe v2 integrations need no other code change — but note that turn prediction, multilingual mode and the keyterm library are not yet in the official SDKs, so those need a direct WebSocket connection for now.
For the full official details, see ElevenLabs’ own blog for the Scribe v3 announcement and ElevenLabs’ pricing page for the current rates, which are the authoritative source if anything here has changed since publication.
Who Should Use ElevenLabs Scribe v3 — and Who Should Wait
Podcasters and YouTubers: the most natural fit. If you already transcribe episodes for show notes, captions or searchable archives, ElevenLabs Scribe v3 is a free accuracy upgrade at the same price, and the keyterm library fixes the recurring frustration of a model mishearing your show’s name or your guests’ names every single episode.
Meeting-heavy teams: the combination of Scribe v3 and the ElevenAI app is worth a trial, especially if your current meeting assistant still produces transcripts you have to correct by hand. Test it on your noisiest real meeting, not a quiet one-to-one, because that is where the accuracy gap shows.
Support and contact-centre teams: Scribe v3 Live plus the Interaction Engine is the pitch here, but be honest about the split — the live model is available now, the engine is a preview. Start with live transcription and Memory in ElevenAgents, and treat the engine as a pilot, not a production dependency, until it is generally available.
Healthcare teams: Scribe v3 Medical is genuinely interesting, but it is a starting point, not a deployment. Run it on your own clinical speech, with your own BAA and your own review process, and measure dosage and drug-name accuracy specifically before it touches a medical record.
Developers building voice agents: the full stack — ElevenLabs Scribe v3 for listening, Eleven Realtime for speaking, the Interaction Engine for orchestration — is the most complete voice-agent package on the market right now. The switching cost is low and the API change is a single model ID, so a side-by-side test against your current stack is cheap.
Who should wait: anyone whose workflow depends on the official SDKs for turn prediction, multilingual mode or the keyterm library — those need a direct WebSocket connection today. And any enterprise that wants a custom model tuned on its own data: that work has not started, and no pricing exists yet, so there is nothing to buy.
ElevenLabs Scribe v3 vs the Competition
Transcription is a crowded market, and ElevenLabs Scribe v3 does not compete in a vacuum. Here is how it lines up against the main alternatives on the things buyers actually compare.
| Tool | Best for | Price signal | What stands out |
|---|---|---|---|
| ElevenLabs Scribe v3 | Meetings, calls, voice agents | $0.20/hour (API) | 102 languages, real-time speaker labels, keyterm library, same price as v2 |
| OpenAI Whisper (API) | General transcription, open weights | $0.36/hour ($0.006/min) | Open-source model you can self-host; strong multilingual baseline |
| Deepgram Nova-3 | Real-time and batch at scale | Usage-based, varies by plan | Long track record in real-time streaming; deep developer tooling |
| AssemblyAI | Developers needing enrichment | Usage-based, varies by plan | Speaker labels, sentiment and topic detection built in |
| Google Chirp 3 / Speech-to-Text | Google Cloud shops | Usage-based, varies by plan | Tight Google Cloud integration; strong on long-form audio |
On price alone, ElevenLabs Scribe v3 undercuts Whisper’s API by roughly 45% per hour, which is a real saving at volume. On features, the differentiators are the real-time speaker labels in Scribe v3 Live and the keyterm library — both aimed at production use rather than demos. Where rivals still lead is in breadth: Deepgram and AssemblyAI have years of developer tooling, streaming infrastructure and enrichment features (sentiment, topic detection, entity extraction) that ElevenLabs is still building out, and Whisper’s open weights let you self-host entirely, which no ElevenLabs model allows.
The deeper comparison is not model against model but stack against stack. ElevenLabs is the only vendor here that also sells the text-to-speech model, the speech-to-speech model, the agent platform and the orchestration layer — and that integration is the point of the Summit package. If you only need transcription, compare on accuracy and price alone. If you are building a voice agent, the question is whether one vendor’s whole stack beats assembling best-of-breed pieces yourself. For a broader view of the transcription market, see our best AI transcription software guide for 2026.
ElevenLabs Scribe v3 Limitations and Honest Caveats
- Vendor benchmarks: the 27% and 43% error-reduction figures are ElevenLabs’ own measurements on its own test sets. No independent evaluation has been published yet. Run your own comparison on your own audio before switching.
- Half the package is not shipping yet: the Interaction Engine is a preview, custom-model tuning has not started, and the financial-services model is only planned. The announcement is a roadmap as much as a release.
- SDK gaps: turn prediction, multilingual mode and the keyterm library are not in the official SDKs yet — they need a direct WebSocket connection, which is more engineering work.
- Add-on pricing: keyterm features are billed on top of the $0.20/$0.22 base rates, so a heavily customised deployment costs more than the headline price suggests.
- Closed models: Scribe v3 weights are proprietary. You cannot self-host or fine-tune them, unlike Whisper. Private deployment is available to enterprise customers, but that is not the same as owning the model.
- Memory is a privacy decision: ElevenAgents Memory only helps if the agent stores what customers said. Decide explicitly what gets remembered, for how long, and how a customer deletes it — before you turn it on.
- Medical use needs your own validation: Scribe v3 Medical is HIPAA-eligible with a BAA, but eligibility is not safety. Test it on your own clinical speech, including dosages, before it touches a patient record.
- Enterprise customisation has no price: custom models tuned on your data are coming, but no pricing, no timeline and no self-serve path have been announced.

ElevenLabs Scribe v3 FAQs
What is ElevenLabs Scribe v3?
ElevenLabs Scribe v3 is ElevenLabs’ third-generation speech-to-text model, launched on 10 November 2026. It transcribes recorded audio in 102 languages, and ElevenLabs says it makes 27% fewer word errors than Scribe v2. It is the default transcription model in the ElevenLabs app and available in the API as scribe_v3.
How much does ElevenLabs Scribe v3 cost?
In the API, ElevenLabs Scribe v3 costs $0.20 per hour of audio — the same as Scribe v2. Scribe v3 Live, the real-time version, costs $0.22 per hour. In the web app, transcription costs 330 credits per minute from your plan’s monthly credit pool. Keyterm features are billed as add-ons.
What is the difference between Scribe v3 and Scribe v3 Live?
ElevenLabs Scribe v3 transcribes recorded audio, such as meetings and podcasts. Scribe v3 Live transcribes audio as it arrives — phone calls, live captions, voice agents — and adds real-time speaker labels, turn prediction and a multilingual mode that handles mid-sentence language switches. ElevenLabs says Live makes 43% fewer errors than Scribe v2 Realtime.
Is ElevenLabs Scribe v3 free?
Partly. The Free plan includes 10,000 monthly credits (about 10 minutes of audio) and access to the Scribe models. Paid plans start at $6 per month for heavier use and commercial licensing. There is no separate free tier for the API beyond the plan credits.
What is the ElevenLabs Interaction Engine?
The Interaction Engine is ElevenLabs’ orchestration layer for voice agents, previewed on 11 November 2026. It assigns the right model to each part of a conversation — cheaper models for routine questions, stronger ones for complex moments — and manages turn-taking and interruptions. It is in preview, not generally available.
Does ElevenLabs Scribe v3 support medical transcription?
Yes, through Scribe v3 Medical, a variant tuned for healthcare terminology that launched alongside the main model. It is HIPAA-eligible with a business associate agreement, but any healthcare team still needs its own BAA, its own review process and its own accuracy testing on real clinical speech before deployment.
How do I switch from Scribe v2 to Scribe v3?
Change the model ID to scribe_v3 (or scribe_v3_live for real-time) in your API integration. No other code change is needed. Note that turn prediction, multilingual mode and the keyterm library are not yet in the official SDKs, so those features need a direct WebSocket connection for now.
Is ElevenLabs Scribe v3 better than Whisper?
On price, yes: $0.20 per hour against Whisper’s $0.36 per hour in the API. On accuracy, ElevenLabs claims a 27% error reduction over its own previous model, but no independent head-to-head against Whisper has been published yet. Whisper’s advantage is that its weights are open and you can self-host it. The honest answer is to run both on your own audio and compare.
ElevenLabs’ New York Summit was not really a launch event for one model. It was a statement that the voice-AI market has moved from selling models to selling the whole conversation: ElevenLabs Scribe v3 to hear, Eleven v4 to speak, Eleven Realtime to converse, the Interaction Engine to orchestrate, and Memory to remember. The models in that stack are real, shipping, and — in Scribe v3’s case — cheaper than the main alternative at the same price as their predecessor. The orchestration and customisation layers are previews and roadmaps, and they should be judged when they ship, not when they are announced.
For most readers, the action is simple and low-risk: try ElevenLabs Scribe v3 on an hour of your own audio this week, compare it with whatever you use today, and let your own transcript decide. If the 27% holds on your speech, the switch costs nothing. If it does not, you have spent twenty cents to find out.
