Grok Voice Think Fast 2.0: AA Speech Lead—and an Aug 5 Default Cutover
xAI’s Grok Voice Think Fast 2.0 cites Artificial Analysis speech-to-speech quality at 82.9% (ahead of GPT-Realtime-2.1 High at 79.1%), 0.70s time-to-first-audio, and $0.08/min. grok-voice-latest migrates to 2.0 on August 5 unless pinned—force a pin strategy before the default flips.
Voice is becoming a primary enterprise surface for support, sales, and agent workflows—not a demo. Shipping a measurable step on third-party speech-to-speech boards and a dated default API migration forces teams to re-evaluate cost, latency, and pin strategy before August 5.
xAI announced Grok Voice Think Fast 2.0, its most capable speech-to-speech voice model. The company primary cites Artificial Analysis leadership on speech-to-speech quality (82.9% vs 79.1% for GPT-Realtime-2.1 High), 0.70s time to first audio (down from 1.25s on 1.0), roughly 0.4× the reasoning tokens of its predecessor in production, and $0.08 per minute of audio. grok-voice-latest migrates to 2.0 on August 5, 2026 unless pinned.
What shipped#
| Item | Detail (x.ai primary) |
|---|---|
| Model | Grok Voice Think Fast 2.0 |
| Focus | Speech reasoning, transcription accuracy, conversational tool use |
| AA Speech-to-Speech Quality | 82.9% (1.0: 75.7%; GPT-Realtime-2.1 High: 79.1%) |
| tau-Voice agentic | 56.5% vs 45.7% for GPT-Realtime-2.1 High (lab table) |
| Latency | 0.70s time to first audio (1.0: 1.25s) |
| Pricing | $0.08 / min of audio |
| Default cutover | grok-voice-latest → 2.0 on Aug 5, 2026 (pin grok-voice-think-fast-1.0 to stay) |
| Production claim | A/B on Starlink support/sales lines: higher conversion and containment |
xAI positions 2.0 as reasoning while speaking—tool calls often fire before the first sentence finishes—without adding turn-taking latency. Transcription claims include 1.5–2.0× lower word-error relative to Deepgram Nova 3 and ElevenLabs Scribe v2 on internal multilingual short-phrase tests, with larger gaps under noise/telephony compression.
Independent rankings#
| Source | Metric | Snapshot |
|---|---|---|
| Artificial Analysis Speech-to-Speech | Quality Index 82.9% | Cited on x.ai launch post as leading GPT-Realtime-2.1 High (79.1%) and Gemini 3.1 Flash High (69.5%) |
| AA / lab tau-Voice | Agentic voice 56.5% | Lab table vs 45.7% (GPT-Realtime-2.1 High) |
Distinct from Grok 4.5 (text/coding flagship), Build Mode, and Imagine media models. Continues the Voice Agent API / Think Fast line (1.0 earlier in 2026). Competes with OpenAI GPT-Live / Realtime and Google Gemini voice stacks.
Limits#
- tau-Voice and WER comparisons vs Deepgram/ElevenLabs are lab tables.
- Starlink A/B conversion/containment is production claim, not a public third-party study.
- Pin
grok-voice-think-fast-1.0before Aug 5 if you need the prior SKU.