
ElevenLabs Scribe v3: Second on AA-WER at 1.95%, Behind a Preview
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 184 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 523 tok/s
- AlibabaNEWQwen: Qwen4 Max2026-11-16$1.50 / $4.50 per 1M tokens · 262 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 94 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 291 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 99 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 132 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 83 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 122 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 104 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 66 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 89 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 41 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 216 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 108 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 160 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
The headline number for ElevenLabs Scribe v3 is 1.95% word error rate, second place on Artificial Analysis' AA-WER v2.2 non-streaming board, behind Alibaba's Fun-Realtime-ASR-preview at 1.73% and ahead of Microsoft AI's MAI-Transcribe-2 at 2.04%. That sentence is accurate and it is also the least useful way to read the result, because 1.95% is not a measurement — it is a weighted average of three different test sets, and on one of them Scribe v3 is not second at all. It is first, ahead of the model that outranks it. ElevenLabs shipped Scribe v3 on 9 November 2026, replaced Scribe v2 in the same API call, and cut the price of every Scribe model on 17 November. Ten days in, the interesting question is no longer whether the model is good. It is which of the three numbers you should believe.
What the 1.95% is actually made of
Artificial Analysis does not publish one WER per model. It publishes a composite (the AA-WER Index) built from three cleaned corpora, and each corpus has a different failure mode: read speech from LibriVox-style audiobooks, parliamentary speech from the European Parliament, and spontaneous multi-speaker earnings calls in accented English. Averaging them into a single index is a reasonable convenience and a bad procurement tool, because the three do not rank the same models the same way.
Here is what the current board shows for the top four, broken out. Every figure below is Artificial Analysis' own, read from the board's data payload rather than the rounded display:
• Composite AA-WER — Fun-Realtime-ASR-preview 1.73%, Scribe v3 1.95%, MAI-Transcribe-2 2.04%, Scribe v2 2.18%.
• The core AA-STT set — Scribe v3 1.47%, Scribe v2 1.51%, MAI-Transcribe-2 1.83%, Fun-Realtime-ASR-preview 1.97%. Scribe v3 wins this one outright, and the model ranked above it finishes last of the four.
• VoxPopuli (parliamentary) — Fun-Realtime-ASR-preview 1.17%, MAI-Transcribe-2 1.50%, Scribe v3 1.52%, Scribe v2 1.74%. Scribe v3 is third here, a hair behind Microsoft.
• Earnings-22 (spontaneous, accented, multi-speaker) — Fun-Realtime-ASR-preview 1.80%, MAI-Transcribe-2 2.99%, Scribe v3 3.34%, Scribe v2 3.95%. This is where the ordering flips hardest: Microsoft beats ElevenLabs by 0.35 points, and the composite's 0.09-point gap disappears.

The pattern is not random. Scribe v3 is strongest on clean read speech and weakest on the corpus that resembles a real meeting — overlapping speakers, disfluencies, telephone bandwidth, accents. That is also the corpus most enterprise buyers actually care about, which is why the composite's ordering and a contact centre's experience may diverge. If your audio is podcasts and studio interviews, the 1.47% figure is closer to your world than the 1.95%. If it is sales calls, the 3.34% is.
The model above it is a preview
Fun-Realtime-ASR-preview holds first place on the composite, and it does so on the strength of the two hardest corpora rather than the easy one. It also runs at a median speed factor of 20.3 — roughly a twentieth of real time — which means an hour of audio takes about three minutes, and it lists at $2.82 per 1,000 minutes. Scribe v3 transcribes at 76.8× real time for $3.33 per 1,000 minutes. So the leader is slower, mostly cheaper, and carries a name that says preview.
None of that makes the ranking wrong; the index measures accuracy, not readiness. It does mean the sentence "Scribe v3 is second" implies a first place held by a production model, and it is not. Alibaba's entry has no release date on the board, no published per-language breakdown, and the word "preview" in the identifier. A team choosing a vendor this quarter is comparing a shipping product against an evaluation build, and the 0.22-point gap between them is smaller than the difference between two corpora inside the same index.
v2 to v3, measured two ways
ElevenLabs' own launch numbers for Scribe v3 are a 27% reduction in word errors against Scribe v2 across 102 languages on an internal FLEURS run — 16.94% down to 12.38% — with improvement claimed in every language tested, plus a 12% reduction in diarization errors and 17% when three or more speakers are present. Those are vendor figures, run on the vendor's own version of a public benchmark, and ElevenLabs is unusually candid that its FLEURS averaging is per-language rather than corpus-level and therefore not comparable to published FLEURS numbers.
Artificial Analysis, running its own harness, shows a smaller gain: 2.18% to 1.95% on the composite, about 11% relative, and 53.5× to 76.8× on throughput. Both statements can be true. A 27% relative improvement on multilingual FLEURS and an 11% relative improvement on three English corpora are measuring different things, and the gap between them is the most useful fact in the launch material — it tells you the win is concentrated outside English read speech, which is exactly where the vendor says it looked.

The feature delta between the two generations is narrower than the marketing implies. Scribe v3 keeps the v2 feature set: up to 32 speakers in diarization, 1,000 keyterms for batch, word-level timestamps, entity detection, transcript editing, export to SRT and DOCX. It adds automatic language inference for a recording, a numbers-format control, secondary language codes, and on the realtime side a 10,000-term keyterm library and turn prediction. It also drops something: there is no verbatim mode. Microsoft's MAI-Transcribe-2 offers both verbatim and clean styles through one switch, and for compliance and QA transcripts that is a functional difference, not a cosmetic one.
What 17 November did to the arithmetic
Eight days after launch, ElevenLabs cut Speech to Text pricing across the board: $0.20 per hour for batch transcription and $0.22 per hour for realtime, on self-serve and Business plans, for every Scribe model. That is $3.33 and $3.67 per 1,000 minutes. It is the number that makes the accuracy table actionable, because at the old rates a 0.09-point accuracy deficit to Microsoft was easy to defend and at the new ones the price gap is the whole argument.

One caution about reading the leaderboard's price column: Scribe v3's $3.33 is exactly $0.20 an hour and Scribe v2's $3.67 is exactly $0.22 an hour, which are the batch and realtime rates in ElevenLabs' current list. Two ElevenLabs rows priced at two different SKUs will read as a 34-cent-per-thousand v2-to-v3 discount that does not exist. Check what the tracker sampled before drawing a pricing conclusion from a ranked table.
Worth stating plainly for anyone building the pipeline behind the transcript: none of these speech models — not Scribe v3, not MAI-Transcribe-2 — sit on OrcaRouter's roster, and the rates above are ElevenLabs' and Microsoft's, not ours. What we route is the layer the transcript feeds: meeting notes to one model, compliance review to another, all behind one key, with provider list prices passed through at 0% markup and automatic failover between providers. That matters more in speech pipelines than elsewhere, because the transcript is the input, not the product, and the product is usually three model calls downstream.
What the next two weeks should settle
Two things are missing from this picture and both are temporary. The first is the streaming board. Scribe v3 Live shipped alongside the batch model with a claimed 31% word-error reduction and 43% lower semantic WER than Scribe v2 Realtime across 102 languages, plus the first realtime diarization ElevenLabs has shipped — 3.1% error on the English slice of a 16-language set against 5.9% to 7.3% for other realtime models. Those are vendor numbers on a vendor-aligned reference, and the streaming leaderboard has not plotted the model yet, so the claim is currently unfalsifiable rather than confirmed.
The second is the medical variant, Scribe v3 Live Medical, which is not an API model at all: it lives in ElevenAgents, is in early access, needs the turn_v3 turn model, has no separate rate, and carries a documentation warning not to send protected health information until HIPAA eligibility is confirmed for it. The batch API medical model remains Scribe v2 Medical. A family launched with a hole where its clinical batch product should be.
So the honest one-line read of Scribe v3 ten days in: it is the best non-streaming model on clean read speech, second-best on meetings, the fastest thing ElevenLabs has shipped, and — after 17 November — cheap enough that the accuracy argument and the price argument finally point the same direction. The only thing standing above it on the board is a preview, and previews do not sign contracts.
