xAI releases Grok Voice Transcribe 2.0 at unchanged prices

xAI has released Grok Voice Transcribe 2.0, a speech-to-text model for prerecorded and live audio. The company kept pricing at $0.10 per audio hour for batch transcription and $0.20 per hour for streaming, with diarization, timestamps, and key-term biasing included. Existing Speech-to-Text API integrations can adopt the model without code changes.

At launch, Transcribe 2.0 ranks first for accuracy among 32 streaming models on the public Artificial Analysis leaderboard. xAI also reports that the model makes roughly half as many errors as Transcribe 1.0 across its internal real-world evaluations.

Hard audio drives the gains

xAI evaluated the model on four internal datasets drawn from production traffic: customer-support calls, conversations with Grok, spoken credentials such as account codes and email addresses, and short multilingual commands. Transcribe 2.0 outperformed its predecessor across all four sets and led every model xAI tested on telephony audio.

Word error rate, or WER, measures inserted, omitted, and incorrectly transcribed words against a reference transcript. Lower scores indicate fewer errors. Differences of a few percentage points can materially affect downstream systems when the audio contains names, identifiers, technical terms, or commands.

Multilingual short phrases produced the largest reported improvement. Transcribe 2.0 supports dozens of languages, detects the language automatically, and handles language changes within one recording. On xAI’s short-phrase benchmark, WER fell from 20.6% to 6.8%, a 67% relative reduction. These tests target brief inputs such as vehicle commands and smart-speaker requests, where the model has little audio available for language detection.

The endpoint bundles production features

The API includes the core controls required for transcription pipelines and real-time voice applications:

• Batch and streaming modes for uploaded files, URLs, and live audio
• Word-level timestamps with confidence scores for individual words
• Speaker diarization that labels speakers at no additional charge
• Multichannel transcription for as many as eight independently processed channels
• Key-term biasing for up to 100 domain-specific terms per request
• Text normalization for numbers, dates, currencies, and email addresses
• Filler-word removal for cleaner transcripts
• Turn detection that helps voice agents identify when a speaker has finished

Pin 1.0 before the default changes

xAI plans to make Transcribe 2.0 the default Speech-to-Text model and deprecate 1.0 in the weeks following the announcement. Unpinned integrations will receive the new model automatically once that change takes effect. The announcement does not specify a final retirement date for 1.0.

Production teams that need a controlled rollout can pin grok-voice-transcribe-1.0 while testing the new release. A representative evaluation should cover domain terminology, noisy audio, overlapping speakers, endpointing latency, diarization quality, and any identifiers that downstream software must reproduce exactly.

Production traffic shaped training

Transcribe 2.0 uses the audio foundation model behind Grok Voice. According to xAI, related systems handle tens of thousands of customer-support calls each day, transcribe millions of hours of video narration, and run the Grok assistant in Tesla vehicles.

xAI says its training data includes noisy, multilingual recordings collected across varied environments, followed by additional post-training. That data mix targets conditions such as 8 kHz call-center audio, compression artifacts, background noise, and overlapping speech, all of which tend to expose weaknesses that clean studio recordings conceal.

Loom feeds transcripts to Cursor

xAI identifies Atlassian as the launch customer for Transcribe 2.0. Atlassian found it more accurate than its previous system and now uses the model to transcribe Loom videos, according to the announcement.

The demonstrated workflow turns a recorded Loom change request into a transcript, passes that text to Cursor, and lets the coding agent apply the requested edits. Transcription errors in variable names, account numbers, or technical instructions can propagate directly into generated code or automated actions, making accuracy on credentials and short commands especially relevant to agent workflows.

The leaderboard lead needs local testing

xAI’s comparisons include ElevenLabs Scribe v2, Deepgram Nova-3, Google Chirp 3, Azure Speech to Text, AssemblyAI Universal-3.5, and OpenAI’s live transcription. Public leaderboard results provide a useful reference, while xAI’s four production datasets remain internal and cannot be independently inspected.

Cross-vendor results also vary with language, acoustic conditions, latency settings, formatting rules, and scoring methods. Developers can evaluate the model in the xAI console and review request parameters, limits, and integration details in the API docs.