Scribe v3 is ElevenLabs’ new batch speech to text model, launched on 10 November 2026 with a realtime sibling, Scribe v3 Live. It costs $0.20 per hour of audio, about $0.0033 a minute, and Scribe v3 Live costs $0.22 an hour, the same rates as Scribe v2. That is 44% under OpenAI’s whisper-1 at $0.006 a minute, a model OpenAI shuts down on 26 February 2027. Keyterm prompting and transcript editing are billed on top, and with both a Scribe v3 hour comes to $0.32, still under Whisper.
What Scribe v3 changes
ElevenLabs kept the Scribe v2 meter and put a better model behind it, so this is a free upgrade for anyone already on the API: switching is a change of model ID to scribe_v3 or scribe_v3_live.
The announcement went out on X on 10 November 2026, with a short video.
Introducing Scribe v3 and Scribe v3 Live: our most accurate Speech to Text models yet.
Built for real-world audio, Scribe excels in situations with extreme background noise, knows pauses from the end of sentences, and keeps track of who said what in a busy meeting.@ElevenLabs
https://x.com/ElevenLabs/status/2120217467100541396
Scribe v3 for recorded audio
Scribe v3 is the batch model: you send a file, you get a transcript back. Its headline gain is speaker labelling (diarization) in meetings with three or more people. The output controls listed in the ElevenLabs launch post are the ones a production pipeline actually needs:
- keyterm prompting with up to 1,000 terms, for brand names, drug names or jargon;
- a choice between numbers written as digits or as words;
- language hints for recordings that may contain several languages;
- transcript edits written in plain English, such as a rule to write every date as YYYY-MM-DD.
Scribe v3 Live for calls and voice agents
Scribe v3 Live is the streaming version, for voice agents and live captions. It brings speaker labels to realtime transcription, and its turn detection reads the conversation instead of waiting on a silence timer, so a caller who pauses mid-number is not cut off. It also adds a multilingual mode and a keyterm library that holds thousands of terms.
The benchmarks ElevenLabs published
These are ElevenLabs’ own numbers, on its internal version of the FLEURS benchmark (102 languages) and a meeting and call dataset it built. Treat them as a vendor claim until an independent leaderboard scores the model.
- Scribe v3 makes 27% fewer word errors than Scribe v2, with German improving 42% and English 23%.
- Scribe v3 Live makes 43% fewer errors, on semantic word error rate, than Scribe v2 Realtime.
- On recordings with three or more speakers, Scribe v3 makes 17% fewer diarization errors than Scribe v2.
- On English calls, Scribe v3 Live has a 3.1% diarization error rate against 5.9% to 7.3% for the other realtime models ElevenLabs tested.
The second post in the launch thread attaches four images with these results.
Scribe v3 has the best transcription and diarization (labeling who said what, when) accuracy of any model we tested.
Scribe v3 Live supports diarization for the first time and makes 43% fewer errors than its predecessor. Turn detection now uses the whole conversation rather than@ElevenLabs
https://x.com/ElevenLabs/status/2120217470112063653
What people are already building with it
Five days after launch, public work on Scribe v3 is thin, and I would rather show three real builds than pad the list.
Eneco’s voice agent in Dutch
Eneco, a Dutch energy supplier with more than 2 million customers, tested Scribe v3 Live on its ElevenAgents phone agent before launch. On real customer calls, the agent got Dutch postcodes right on the first try 40% more often and house numbers 15% more often. The figures come from ElevenLabs’ launch post, a showcase rather than an audit, but they measure what matters on a phone line: digits.
A Claude Code skill that writes SRT subtitles
The transcribe skill by shahafg1 turns audio or video into SRT subtitles with Scribe v3, keeps Scribe v2 as a fallback and is tuned for Hebrew and other right-to-left languages. It passes up to 1,000 keyterms and labels up to 32 speakers per cue. One correction to its README: keyterms do not add “about 20%” to the bill, they add 25% ($0.05 on a $0.20 hour).
A converter from diarized JSON to subtitles
BlohoJo’s Diarized JSON SRT Converter reads the per-word, speaker-labelled JSON that Scribe v3 returns and turns it into broadcast-style subtitles: dialog dashes, balanced line splits, reading-speed limits and sound tags such as [laughs] kept or dropped with one switch. It is plain Python with a Windows interface.
Scribe v3 pricing, base rate and add-ons
ElevenLabs bills the Speech to Text API in US dollars per hour of audio, metered by the audio minute, so silence in the file is billed like speech. The ElevenLabs API pricing page groups Scribe v3 with Scribe v2 and Scribe v2 Medical on one meter, and Scribe v3 Live with Scribe v2 Realtime on another.
The ElevenLabs API meter, billed per hour of audio
The batch meter
On 15 November 2026 the grid reads the same in every plan column, so the plan you pick changes how many hours are included, not the rate.
| Scribe v3 batch | Price per hour of audio |
|---|---|
| Base transcription | $0.20 |
| Keyterm prompting | +$0.05 |
| Transcript editing | +$0.07 |
| Hours included, smallest to largest plan | 4 h 30 min to 4,950 h |
The Live meter
Streaming costs 10% more at the base, but the add-ons cost far more. Keyterm prompting goes from $0.05 to $0.08 an hour, transcript editing from $0.07 to $0.12, and the keyterm library, which only Scribe v3 Live offers, adds another $0.08. Included hours run from 2 h 30 min to 4,500 h depending on the plan.
What the add-ons do to an hour
The base rate is the price only if you send raw audio and take whatever comes back. Each add-on is a fixed amount per hour of audio, so the percentages below hold at any volume:
- Scribe v3 with keyterms: $0.25 an hour, 25% above the base.
- Scribe v3 with keyterms and transcript editing: $0.32 an hour, 60% above the base.
- Scribe v3 Live with keyterms: $0.30 an hour.
- Scribe v3 Live with keyterms, keyterm library and transcript editing: $0.50 an hour, more than double the base.
A fully loaded Scribe v3 Live hour costs $0.50, against $0.22 at the base rate. On a voice agent, the add-ons cost more than the transcription itself.
At 1,000 hours a month: $200 for plain Scribe v3, $320 with both batch add-ons, $500 for Live with everything on.
Scribe v3 against Whisper and the other transcription APIs
Every rate below was read on the vendor’s own pricing page or model page on 15 November 2026, converted to an hour of audio. All are batch rates unless the row says otherwise.
| Model | Price per hour of audio |
|---|---|
| xAI Speech to Text (REST) | $0.10 |
| DeepSeek Speech to Text (ASR), list price | $0.116 |
| ElevenLabs Scribe v3 | $0.20 |
| Mistral Voxtral Mini Transcribe 3 | $0.21 |
| Mistral Voxtral Small Transcribe 3 | $0.24 |
| OpenAI gpt-transcribe | $0.27 |
| OpenAI whisper-1 (deprecated) | $0.36 |
| ElevenLabs Scribe v3 Live (streaming) | $0.22 |
| OpenAI gpt-live-transcribe (streaming) | $1.02 |
Against OpenAI, per minute
OpenAI prices by the minute, so here is the same comparison in its unit. whisper-1 costs $0.006 a minute, gpt-transcribe $0.0045 and Scribe v3 about $0.0033. On 1,000 hours that is $360, $270 and $200.
whisper-1 is also on its way out: OpenAI’s deprecations page lists its shutdown for 26 February 2027, with gpt-transcribe or gpt-live-transcribe as replacements (the full lineup is in the Whisper API pricing breakdown). If you must touch that integration anyway, Scribe v3 with keyterms is still cheaper than both.
On streaming the gap is wider still: gpt-live-transcribe costs $0.017 a minute, $1.02 an hour, almost five times Scribe v3 Live.
Against Voxtral 3, xAI and DeepSeek
Scribe v3 is not the cheapest hour on the market. Three providers sit close to it or under it:
- Mistral’s Voxtral Mini Transcribe 3 costs $0.0035 a minute, $0.21 an hour, a cent above Scribe v3; its realtime version costs $0.48 an hour. The full grid is in the Voxtral 3 transcription pricing page.
- xAI’s Speech to Text costs $0.10 an hour over REST and $0.20 streaming, half of Scribe v3 in batch and 2 cents under Scribe v3 Live.
- DeepSeek’s Speech to Text (ASR) lists at $0.116 an hour and currently shows a limited-time price of $0.058, half off, with no end date on the page.
Before switching for price, check that the cheaper meter labels speakers and accepts a term list as long as yours. If a person reads the transcript, those two features are worth more than the dime.
Which hour I would buy
For meeting notes, interviews and subtitles, I would buy Scribe v3 with keyterms at $0.25 an hour: it is still cheaper than whisper-1 and gpt-transcribe, and the keyterm list fixes the names and acronyms that cost the most to correct by hand. Transcript editing at $0.07 is worth it only if you would otherwise run the transcript through a text model to reformat it.
For a voice agent, start with plain Scribe v3 Live at $0.22 and add the keyterm library only when the vocabulary outgrows a list, because the full Live stack more than doubles the hour. If you already pay ElevenLabs for speech, the ElevenLabs cost per hour page covers the text to speech side of the same bill.
For bulk archives nobody reads line by line, I would not buy Scribe v3 at all: xAI at $0.10 does that job for half the money.