Meta's open-weights Muse model, released 15 October 2026: 816B parameters with 42B active, a 1M-token context, text, image, video and audio input, and a licence that is free below a $20M service threshold.
Scribe v3
New· 6ElevenLabs' speech-to-text models from 10 November 2026: Scribe v3 for recordings across 102 languages and Scribe v3 Live for realtime, which adds speaker labels and mid-sentence language switching.
- Artificial Intelligence - AI
- Speech Recognition
TrustList summary
Not yet independently verified. ElevenLabs' announcement stated no price, but its API pricing page now lists Scribe v3 at $0.20 per hour of audio (read 16 November 2026). The error-rate comparisons are still ElevenLabs' own, run on its own versions of the test sets, and we have not found an independent test. We will update this when it can be confirmed, and remove this note.
Speaker labels now work in realtime with ElevenLabs. Scribe v3 Live, released on 10 November 2026 with the batch model Scribe v3, is the first of the company's live transcription models to separate speakers as it goes, and it adds a multilingual mode for conversations that switch language mid-sentence. The API identifiers are scribe_v3 and scribe_v3_live, and teams on Scribe v2 move by changing the model ID.
Two models for two jobs
Scribe v3 works on recorded audio and is evaluated across 102 languages. It can infer the likely languages in a recording, write numbers as digits or words, take up to 1,000 key terms as prompts and apply formatting rules written in plain English. Scribe v3 Live is the low-latency model: it detects turns from conversation context as well as silence, through a turn_prediction commit strategy, and draws on a key-term library for vocabularies too large for a list. A version tuned for medical conversations runs inside ElevenAgents.
What ElevenLabs says it measured
The company reports 27% fewer word errors than Scribe v2 for the batch model and 43% fewer semantic errors for the live model against Scribe v2 Realtime, both on its internal version of the FLEURS benchmark of short, clean read speech. For speaker separation it reports 12% fewer errors overall, and a 3.1% diarization error rate on the English part of its own 16-language dataset, which it says it will publish. Eneco, the Dutch energy supplier, reports postcodes recognised correctly at first attempt 40% more often in its voice agent. Ruud Huigsloot, Senior Product Manager at Eneco, said: "It's great to see significant results this early, and we are excited to continue collaborating."
Residency and compliance
Processing runs in the US by default, with isolated EU, India and Singapore environments for enterprise customers and private deployment available. ElevenLabs lists a Zero Retention Mode that deletes audio and text when a request completes, SOC 2, ISO 27001 and PCI DSS Level 1, and HIPAA eligibility under a Business Associate Agreement.
Key facts
Maker: ElevenLabs Released: 10 November 2026 Model codes: scribe_v3 (batch), scribe_v3_live (realtime) Languages: evaluated across 102 Default region: US; EU, India and Singapore for enterprise Availability: ElevenAPI, ElevenAgents and the ElevenLabs app
Benchmarks & AI stats
| Benchmark | Official | Community avg |
|---|---|---|
| No benchmark scores yet. Be the first to add one. | ||
“Official” values are editor-approved and feed the ranking. “Community avg” is the mean of member submissions (shown for transparency; it never affects the ranking until an editor approves a value).
Sign in to add a benchmark score for this model.
Scribe v3 launches
Categories & features
- Artificial Intelligence - AI
- Speech Recognition
Trust Score
An earned signal from verification, reviews, awards, transparency and engagement - the vendor can't buy it.
Updated 11/17/2026
Request a demo or quote from Scribe v3
Reviews
Write the first review of Scribe v3
Used it? Your experience helps other buyers decide.
Write a reviewQuestions & answers
No questions yet. Be the first to ask about Scribe v3.
Scribe v3 alternatives
Similar listings buyers compare against this one.
IBM's modular Granite model, released 14 October 2026 in 8B and 30B sizes: one endpoint for chat, scored multiple-choice decisions, safety checks and English speech-to-text, under Apache 2.0.
Google's classification model from 11 November 2026 answers up to 1,000 yes/no, choice or rubric questions over one input of up to 256K tokens, billed at $0.10 per million input tokens with output free.
A two-week preview of Alibaba's next Qwen flagship, open from 11 to 25 November 2026 (UTC+8) through the Token Plan only, with a 1M-token context and 393,216 tokens of output shared with thinking.
The gated twin of Claude Fable 5.2 without its cyber and biology safeguards, released 9 November 2026 for approved security and life sciences organisations at $8 in and $40 out per million tokens.
Open-weight agentic coding model from IQuestLab with a 524,288-token context, for teams that self-host
Anthropic's top general-access model from 9 November 2026: a 1M-token context, 128K output, and $8 in and $40 out per million tokens, below the $10 and $50 of Fable 5.1, with cyber and biology safeguards.
OpenAI's restricted security model from 5 November 2026, for investigating vulnerabilities and verifying fixes: a 1,050,000-token context at $20 in and $100 out per million tokens, for Daybreak Red organisations with further approval.
More on TrustList
Everything here links back to the same verified catalogue. Pick your next stop.
- More Artificial Intelligence - AIBack to the ranked category
- Write a reviewHelp other buyers decide
- CompaniesAgencies, consultancies and IT service providers, ranked by verified reviews.
- ProductsSoftware and SaaS with pricing, features, integrations and alternatives.
- ArticlesGuides, comparisons and research from the editorial desk and the community.
- AwardsAnnual recognition decided by verified reviews and an independent jury.
- LaunchesNew products and releases, voted up by the community every day.
- RequestsBuyers describe what they need; vendors respond directly.
- PeopleReviewers, authors and makers with public profiles.
- ComparePut up to four listings side by side before you shortlist.