ElevenLabs has released Scribe v3, which its model documentation calls its “most accurate speech recognition model for recorded audio”. On the FLEURS benchmark across 102 languages, the documentation says, Scribe v3 lowers the word error rate from 16.94% to 12.38%, “27% fewer word errors than Scribe v2”.
Its API pricing page lists Scribe v3 at $0.20 per hour of audio, the same price it lists for Scribe v2. The real-time version, Scribe v3 Live, is listed at $0.22 per hour, the same as Scribe v2 Realtime. Fewer errors at an unchanged price, by our reading of the two pages.
Business Pill · The Model and the Harness
An evergreen Business Engineer explainer on how models and the harness around them split the value. It teaches the general idea only and does not discuss Scribe v3.
The key insight: As we read it, ElevenLabs has improved its transcription models without raising their list prices: its documentation reports 27% fewer word errors for Scribe v3 than Scribe v2 on FLEURS, while its pricing page lists both at $0.20 per hour, and Scribe v3 Live at the same $0.22 per hour as Scribe v2 Realtime.
The documentation says Scribe v3 uses the same Speech to Text API as Scribe v2: developers change the model_id from scribe_v2 to scribe_v3, and “Existing request parameters work without changes”. It says “Scribe v2 remains available”.
The documentation lists meeting and note-taking tools, media subtitles, recorded lectures and contact-centre call analytics as target uses. It adds a numbers_format parameter, which writes numbers as digits or words, and a secondary_language_codes parameter for mixed-language audio. Scribe v3 keeps support for 90+ languages and speaker diarization for up to 32 speakers, the documentation says.
Scribe v3 Live is the successor to Scribe v2 Realtime, the documentation says. On its speech-to-text page, ElevenLabs says the model works “at a median latency of 150 ms” and is “Built for voice agents, meetings and live captions”.
Word error rate for the Scribe v2 and Scribe v3 generations on three benchmarks, from ElevenLabs model documentation read on 11 November 2026. The Live rows compare Scribe v3 Live with Scribe v2 Realtime. Percentage cuts for the Live rows are our arithmetic.
The Real-Time Numbers
For Scribe v3 Live, the documentation says word error rate on Earnings-22, a set of accented English financial calls, drops from 5.99% to 3.91%. That is about 35% fewer errors, by our arithmetic. On VoxPopuli, European Parliament speech, it says the rate falls from 2.13% to 1.71%, about 20% fewer by our arithmetic.
Scribe v3 Live adds turn prediction, which the documentation says commits a segment “when the model predicts that the speaker has finished their turn, instead of waiting for silence alone”. It also adds a multilingual mode and a keyterm library of up to 10,000 terms. For ElevenLabs’ earlier real-time push, see our piece on its OpenAI-compatible voice model.
ElevenLabs’ speech-to-text page compares Scribe v3 Live’s semantic word error rate with rival real-time models on the languages each supports. It lists 4.8% against 8.3% for Gemini 3.5 Transcribe Live across 75 languages, 3.9% against 6.7% for GPT-Live-Transcribe across 57, and 3.6% against 12.5% for Deepgram Nova-3 across 53.
Our Numbers
At the listed $0.20 per hour, transcribing 1,000 hours of recorded audio with Scribe v3 costs $200, by our arithmetic, the same as with Scribe v2. The real-time tier costs 10% more per hour than the batch tier, $0.22 against $0.20, by our arithmetic.
The Scribe v3 gain is 4.56 percentage points of word error rate on FLEURS (16.94% minus 12.38%), by our arithmetic, which is the 27% cut ElevenLabs reports. For how ElevenLabs ranked on independent voice leaderboards, see our piece on the new voice boards.
The Structural Read
As we read the pages, the business lever is accuracy per dollar: with the hourly price unchanged, the documentation’s drop from 16.94% to 12.38% word error rate is a better product at the same cost to the developer.
As we read the speech-to-text page, the real-time tier is pitched at voice agents: it lists turn prediction, a median latency of 150 ms, and rival comparisons against Gemini 3.5 Transcribe Live, GPT-Live-Transcribe and Deepgram Nova-3, all on ElevenLabs’ own tests.
“Existing request parameters work without changes”
Three Implications
FOR DEVELOPERS ON SCRIBE V2 The documentation says the switch is a change of model_id from scribe_v2 to scribe_v3, and that Scribe v2 remains available.
FOR VOICE-AGENT BUILDERS The documentation says Scribe v3 Live commits a turn when the model predicts the speaker has finished, instead of waiting for silence alone.
FOR BUYERS COMPARING VENDORS The documentation says its FLEURS figures are averaged per clip and per language and are not comparable with corpus-level WER computed another way.
The framework argues: “The agentic surface — the harness — is becoming where value pools in the entire AI stack.”
As we read it, Scribe v3 Live is a component sold into that surface: the documentation lists voice agents and contact centers first among its uses. The framework predates Scribe v3 and does not discuss it.
What Is Not Established
The documentation says ElevenLabs computes FLEURS word error rate per clip and then averages per language, which “produces higher figures than the corpus-level WER that is commonly published, so the numbers are not comparable with results computed another way”. The rival comparisons on the speech-to-text page are ElevenLabs’ own tests; we did not find independent results in the pages we read.
The pricing page says Scribe v3 Live has “43% fewer errors than v2 Realtime”, while the documentation says it has a 44% lower semantic word error rate on FLEURS across 73 languages; the two pages use different wording. The pages we read give no release date and no figures on customers or usage.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
ElevenLabs lists Scribe v3 at $0.20 per hour and Scribe v3 Live at $0.22 per hour, the same prices as the Scribe v2 models they succeed. Its documentation reports 27% fewer word errors for Scribe v3 on FLEURS across 102 languages, and a drop from 5.99% to 3.91% on accented financial calls for Scribe v3 Live.
A note on sourcing. We read ElevenLabs’ model documentation for Scribe v3 and Scribe v3 Live, its API pricing page and its speech-to-text page in full on 11 November 2026. The benchmark figures are ElevenLabs’ own and were not independently tested by us. Differences, percentage changes and costs are our arithmetic. ElevenLabs was not contacted for this piece.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
