# Grok Voice Think Fast 2.0 Ships: 82.9% Score at $0.08/Min

SpaceXAI released Grok Voice Think Fast 2.0, an upgraded speech-to-speech model with an 82.9% AA Quality Index, 0.70s time-to-first-audio, and enterprise pricing at $0.08 per minute. The Grok-Voice-Latest endpoint transitions from v1.0 to v2.0 on August 5.

By Neomanex. Updated 2026-10-07.

SpaceXAI's Grok Voice Think Fast 2.0 is the new speech-to-speech benchmark leader — 82.9% on Artificial Analysis' Quality Index, 0.70s time-to-first-audio, and a 7.2-point gain over v1.0. At $0.08/minute, it's competitively priced against a voice model field that's iterating faster than text models did at the same stage. The August 5 auto-migration of `grok-voice-latest` means every Grok Voice API user gets the upgrade with zero config changes — but it also means if you haven't tested v2.0 yet, you have six days.

## What Happened

SpaceXAI announced Grok Voice Think Fast 2.0 on July 29 via the [x.ai blog](https://x.ai/blog/grok-voice-think-fast-2). The speech-to-speech model delivers across four dimensions:

**Benchmark leadership.** On Artificial Analysis' Speech-to-Speech Quality Index, Grok Voice Think Fast 2.0 scores 82.9% — ahead of GPT-Realtime-2.1 (79.1%) and Gemini 3.1 Flash (69.5%). Specific highlights:

| Metric | Grok Voice 2.0 | Grok Voice 1.0 | GPT-Realtime-2.1 | Gemini 3.1 Flash |
|---|---|---|---|---|
| AA Quality Index | **82.9%** | 75.7% | 79.1% | 69.5% |
| Conversational Dynamics | **95.1%** | 77.8% | 95.7% | 74.3% |
| Agentic Performance (τ-voice) | **56.5%** | 52.1% | 45.7% | 37.7% |
| Time to First Audio | **0.70s** | 1.25s | — | 2.98s |

**Transcription accuracy.** Outperforms dedicated transcription models Deepgram Nova 3 and ElevenLabs Scribe v2 by 1.5–2.0× across 24 languages. The gap widens to ~10× in noisy, real-world settings with telephony compression — this is a speech-to-speech model that handles transcription better than purpose-built speech-to-text tools.

**Reasoning efficiency.** Grok Voice Think Fast 2.0 uses 0.4× the reasoning tokens of v1.0 while reasoning in parallel with speech. Tool calls execute before the end of the agent's first sentence — snappier, cheaper, same intelligence.

**August 5 auto-migration.** The `grok-voice-latest` endpoint transitions from `grok-voice-think-fast-1.0` to `grok-voice-think-fast-2.0` on August 5, 2026. No action needed to upgrade. To stay on v1.0, pin `grok-voice-think-fast-1.0` before then.

**$0.08/minute pricing.** Enterprise API pricing with no subscription gate — available now via the x.ai console.

## Why It Matters

The voice model category is where text models were in early 2025: rapid iteration, aggressive pricing, and no entrenched winner. Grok Voice 2.0 at $0.08/minute with API-accessible full-duplex capability is a direct competitive shot across the bow of OpenAI's voice models — which currently gate real-time voice behind ChatGPT subscriptions with no standalone API pricing for full-duplex.

For developers building voice agents, meeting assistants, or real-time conversational AI: the model landscape is fragmenting fast. You now have GPT-Live (ChatGPT-only, no API), GPT-Transcribe (batch, API), GPT-Live-Transcribe (streaming, API), and Grok Voice 2.0 (full-duplex, API, $0.08/min). Each has different latency, pricing, and API-access tradeoffs. The right choice depends heavily on your use case — and the landscape will look different by October.

SpaceXAI is also using Grok Voice 2.0 in production on Starlink's customer support line ([+1 888 GO STARLINK](tel:+18884678275)), where A/B testing showed significant gains in sales conversion and support containment rates. That's a real-world validation signal — not just benchmarks.

## What Changes for You

1. **Test v2.0 before August 5.** If you're on `grok-voice-latest`, the cutover is automatic. v2.0's faster response time and shorter sentences change conversation pacing — test your integration before the endpoint switches.

2. **Benchmark against your voice use case, not just leaderboards.** The 0.70s time-to-first-audio and 56.5% agentic performance score matter more for real-time voice agents than the overall quality index. Run your own latency and task-completion tests.

3. **Don't lock into a single voice model yet.** The category is evolving faster than text models did at the same stage. Build your voice pipeline with model-swappable architecture — the leader in August won't necessarily be the leader in October.

Canonical: https://neomanex.com/news/grok-voice-think-fast-2