AI tools · September 21, 2026

xAI’s Grok Voice Transcribe 2.0 doubles speech accuracy at the same $0.10/hour price

xAI released Grok Voice Transcribe 2.0 on September 18, and it’s already the default: when you don’t specify a model, the speech-to-text API now routes to the new version (pin `grok-voice-transcribe-1.0` to keep the old one). The headline is accuracy-per-dollar — xAI says the new model delivers twice the accuracy of 1.0 at identical pricing: $0.10 per audio hour for batch, $0.20 per audio hour for streaming, with speaker diarization, word-level timestamps, and key terms included.

The vendor-reported numbers: short-phrase word-error rate across 19 languages falling from 20.6% to 6.8%, and a claimed #1 ranking among 32 streaming models on Artificial Analysis’s AA-WER Streaming leaderboard. Neither figure has been independently reproduced, so treat them as xAI’s claims, not third-party verification. The spec sheet is concrete, though: word-level timestamps with confidence scores, free speaker diarization, up to 8 independent channels, up to 100 key terms per request (50 characters max each), filler-word removal on by default, Smart Turn end-of-turn detection, 12 audio formats, 500MB maximum batch files, and WebSocket streaming at wss://api.x.ai/v1/stt.

It’s the same audio foundation behind Grok Voice, which xAI says handles tens of thousands of customer-support calls daily and runs in Tesla vehicles. It’s already in production at Atlassian, where Loom now transcribes every video with Transcribe 2.0. If you’re building voice agents, meeting notetakers, or transcription pipelines, this is the cheapest way to test the claimed 2× accuracy jump against your own evals.

Why it matters

Speech-to-text is the cheapest input to make voice agents and transcription tools genuinely good, and when the price holds at ten cents an hour while claimed accuracy doubles, every existing voice pipeline should re-run its evals. Independent verification will decide whether the hype holds — but the API is live now, so there’s no waiting to check.

Key facts

Released September 18, 2026; model ID `grok-voice-transcribe-2.0`; now the default when `model` is omitted (pin `grok-voice-transcribe-1.0` for the old model)
Same pricing as 1.0: $0.10 per audio hour batch, $0.20 per audio hour streaming — diarization, word-level timestamps, and key terms included
xAI claims 2× better accuracy than 1.0; short-phrase word-error rate across 19 languages down from 20.6% to 6.8%; vendor-reported #1 of 32 streaming models on the Artificial Analysis AA-WER leaderboard (not independently reproduced)
Features: word-level timestamps with confidence scores, free speaker diarization, up to 8 independent channels, up to 100 key terms per request (≤50 chars), filler-word removal by default, Smart Turn end-of-turn detection
Specs: 12 audio formats, 500MB maximum batch files, WebSocket streaming at wss://api.x.ai/v1/stt
Same audio foundation as Grok Voice — tens of thousands of customer-support calls daily per xAI, also runs in Tesla vehicles; Atlassian’s Loom now transcribes every video with it

Sources

Spot an error? We correct quickly and note it. Contact us via the contact page.

← More AI news