What is Grok Voice Transcribe 2.0?

Grok Voice Transcribe 2.0 is SpaceXAI's speech-to-text model, released September 18, 2026 and available only through the api.x.ai Speech-to-Text API. It ranks first for accuracy among 32 streaming models on the Artificial Analysis leaderboard, roughly doubles the accuracy of Transcribe 1.0, and keeps the same price: $0.10 per hour of audio for batch and $0.20 per hour for streaming, with speaker diarization, word-level timestamps and key-term biasing included. There are no open weights. Transcribe 1.0 will be deprecated in the coming weeks.

Which speech-to-text API is cheapest in 2026?

As of September 20, 2026, batch transcription per hour of audio: Grok Voice Transcribe 2.0 $0.10, Meta Muse Voice Transcribe $0.18, AssemblyAI Universal-3.5 Pro $0.21, ElevenLabs Scribe v2 $0.22, Deepgram Nova-3 $0.26, OpenAI GPT Transcribe $0.27 ($0.0045/min), OpenAI Whisper API $0.36 ($0.006/min). Streaming: Grok $0.20, Deepgram $0.29, ElevenLabs $0.39, AssemblyAI $0.45. Grok is the cheapest on both, and it bundles diarization, which AssemblyAI bills as an add-on.

Which speech-to-text model is most accurate?

On Artificial Analysis's public AA-WER leaderboard, Grok Voice Transcribe 2.0 ranks first among 32 streaming models as of September 18, 2026, and ElevenLabs Scribe v2 sits around a 2.3% word error rate on the non-streaming index. SpaceXAI's internal sets show Transcribe 2.0 cutting short-phrase multilingual WER from 20.6% to 6.8% and telephony WER on the first final transcript from 3.9% to 2.7%. Accuracy differs by audio type, so test on your own recordings: AssemblyAI's own comparisons show Universal-3.5 Pro ahead on code-switching and entity accuracy.

Should I switch from Whisper to Grok Voice Transcribe 2.0?

If you use the hosted Whisper API, yes, on cost alone: $0.36 per hour versus $0.10, with diarization and timestamps included. If you self-host Whisper large-v3 because audio cannot leave your machines, Grok is not a substitute — it is API-only with no open weights. For that case a local Whisper fork or an open model remains the right choice, and the 2026 alternative is Alibaba's Qwen3.8-Omni-Flash if you can use a hosted omnimodal model instead.

Quick Answer

Grok Transcribe 2.0 vs Deepgram vs AssemblyAI vs Scribe v2

Published: September 20, 2026

The comparison at a glance

	Grok Voice Transcribe 2.0	Deepgram Nova-3	AssemblyAI Universal-3.5 Pro	ElevenLabs Scribe v2
Maker	SpaceXAI (formerly xAI)	Deepgram	AssemblyAI	ElevenLabs
Released	September 18, 2026	2025, updated 2026	2026	March 2026 (Realtime later)
Batch price / hour	$0.10	$0.26	$0.21	$0.22
Streaming price / hour	$0.20	$0.29	$0.45	$0.39
Diarization	Included	Included	+$0.02/hr async, +$0.12/hr realtime	Included, up to 48 speakers
Languages	Dozens, auto-detect, mid-recording switching	Multilingual	18 with code-switching (Pro)	90+
Multichannel	Up to 8 channels	Yes	Yes	—
Open weights	No	No	No	No
AA leaderboard	#1 of 32 streaming models	Mid-pack	Strong on entities	~2.3% AA-WER non-streaming
Best for	Cheapest production audio, telephony, voice agents	Low-latency streaming, custom models	Audio intelligence add-ons, medical mode	Multilingual batch, non-speech events

Verified September 20, 2026 against SpaceXAI’s announcement, vendor pricing pages and Artificial Analysis. Prices are pay-as-you-go list rates; volume and annual tiers are lower.

Grok Voice Transcribe 2.0 — what changed

SpaceXAI released Grok Voice Transcribe 2.0 on September 18, 2026, built on the audio foundation model behind Grok Voice — the same stack that runs the Grok assistant in Tesla vehicles and, the company says, tens of thousands of customer-support calls a day. Three facts matter for buyers:

Accuracy roughly doubled versus 1.0. On the public Artificial Analysis leaderboard it ranks first for accuracy among 32 streaming models. SpaceXAI’s four internal production sets all improved: short multilingual phrases went from 20.6% to 6.8% WER, telephony first-final transcripts from 3.9% to 2.7%, and spoken credentials such as phone numbers and email addresses were the third set to improve.
Price did not move. Batch stays at $0.10 per hour of audio, streaming at $0.20 per hour, and diarization, word-level timestamps with confidence scores, and up to 100 key terms per request are included rather than billed as add-ons.
It is API-only. Endpoint api.x.ai/v1/stt, with batch and streaming modes, up to 8 independent channels, filler-word removal, written-form formatting of numbers and dates, and smart turn detection for voice agents. There are no open weights. Existing integrations get the upgrade with no code changes; pin grok-voice-transcribe-1.0 if you need the old model during the deprecation window.

Atlassian is the launch customer: Loom now uses Transcribe 2.0 for screen-recording transcripts and pipes them into Cursor for code changes.

Deepgram Nova-3 — the streaming incumbent

Deepgram’s Nova-3 is the model most voice-agent stacks were built on, with sub-300 ms streaming latency and custom-model training. Pay-as-you-go pricing verified September 8, 2026 is $0.26 per hour batch and $0.29 per hour streaming ($0.0043/min), with $200 of free credit for new accounts. Deepgram’s advantage is maturity — keyword boosting, custom vocabularies, on-prem deployment for enterprise — not price: Grok now undercuts it by 2.6x on batch.

AssemblyAI Universal-3.5 Pro — the feature stack

AssemblyAI’s Universal-3.5 Pro is priced at $0.21 per hour async and $0.45 per hour realtime, billed per second with no minimums. Diarization is an add-on (+$0.02/hr async, +$0.12/hr realtime), as are keyterm prompting and Medical Mode (+$0.15/hr). The July 14, 2026 Sync API returns finished transcripts for short clips at the realtime rate without polling. AssemblyAI’s own head-to-heads claim wins on code-switching across its 18 supported languages and on entity error rate, which is the metric that matters for names, drugs and account numbers. Its Voice Agent API bundles STT, LLM and TTS on one WebSocket at $4.50/hr.

ElevenLabs Scribe v2 — multilingual batch

Scribe v2 costs $0.22 per hour batch and $0.39 per hour realtime ($0.28 on annual Business plans), plus $0.05/hr for keyterm prompting and $0.07/hr for entity detection. It supports 90+ languages, diarization for up to 48 speakers, non-speech event tagging and ~150 ms realtime latency. On Artificial Analysis’s non-streaming index it reports around 2.3% AA-WER, which puts it close to Grok on clean audio; the gap opens on noisy telephony, where SpaceXAI trained specifically.

The other two you will be quoted

Meta Muse Voice Transcribe launched September 1, 2026 at $0.18 per hour for both streaming and batch, with real-time diarization for 20+ speakers included and a zero-data-retention option at the same price. It undercut everyone for 17 days until Grok 2.0 shipped at $0.10.
OpenAI GPT Transcribe is $0.0045/min ($0.27/hr); the legacy Whisper API is $0.006/min ($0.36/hr). Whisper large-v3 remains the default for self-hosting because it is open; none of the four models above are.

Which one to pick

Lowest bill, hardest audio (call centres, in-car, IVR): Grok Voice Transcribe 2.0. It leads the streaming leaderboard and is the cheapest by a wide margin.
Existing Deepgram voice agent with custom models: stay unless the audio bill is material; the migration cost outweighs $0.16/hr on small volumes.
Medical, legal or entity-heavy transcripts with downstream summarisation: AssemblyAI, for the audio-intelligence add-ons and code-switching.
Podcasts and video in many languages: ElevenLabs Scribe v2 or Grok — run both on a 10-file sample; they trade wins by language.
Audio that cannot leave your infrastructure: self-host Whisper or an open model. Nothing in this table gives you weights.