Skip to content
NewsModel release

xAI releases Grok Voice Think Fast 2.0, outscoring GPT-Realtime-2.1 on Artificial Analysis benchmarks

· by Pondero Newsdesk

The short version

xAI launched Grok Voice Think Fast 2.0 on July 29, 2026, routing the grok-voice-latest alias to the new model on August 5. Artificial Analysis benchmarks show it at 82.9% on the Speech-to-Speech Quality Index, ahead of GPT-Realtime-2.1 at 79.1% and Gemini 3.1 Flash at 69.5%.

xAI releases Grok Voice Think Fast 2.0, outscoring GPT-Realtime-2.1 on Artificial Analysis benchmarks

Grok Voice Think Fast 2.0 cut time to first audio from 1.25 seconds to 0.70 seconds and scored 82.9% on Artificial Analysis's Speech-to-Speech Quality Index, ahead of GPT-Realtime-2.1 at 79.1% and Gemini 3.1 Flash at 69.5%, per xAI's announcement. xAI launched the model on July 29, 2026, and routed the grok-voice-latest API alias to it on August 5.

What

Grok Voice Think Fast 2.0 is now the model behind the grok-voice-latest alias, per xAI's speech-to-speech documentation. Developers using the grok-voice-latest endpoint had until August 5 to pin grok-voice-think-fast-1.0 to avoid the automatic upgrade. Those who did not pin migrated to version 2.0 on that date with no code changes required.

Time to first audio dropped from 1.25 seconds to 0.70 seconds relative to version 1.0, per xAI's announcement. Transcription accuracy improved 1.4x over grok-voice-think-fast-1.0 across 24 languages, and 1.5 to 2.0 times relative to Deepgram Nova 3 and ElevenLabs Scribe v2, per the same evaluation. Conversational Dynamics Full Duplex Bench score reached 95.1% against 77.8% for version 1.0. Agentic Performance on the tau-voice benchmark reached 56.5%, up from 52.1%.

Reasoning efficiency also changed. Per xAI, Think Fast 2.0 uses roughly 0.4x the reasoning tokens per response that version 1.0 used. In production, per the announcement, this results in tool calls completing before the end of the agent's first spoken sentence rather than causing a noticeable pause. Pricing is $0.08 per minute of audio, per xAI's pricing page.

xAI disclosed that it A/B tested the model on the Starlink customer support line before launch. Per the blog post, the test showed increased sales conversion and support containment rates versus version 1.0. Those are vendor self-claims not independently audited.

Why it matters

Two numbers make this release notable for developers actively evaluating real-time voice APIs. First, the 82.9% quality index puts Think Fast 2.0 ahead of GPT-Realtime-2.1 on a benchmark run by an independent measurement firm rather than one of the model vendors. Second, the latency gap is large enough to matter in telephony: 0.70 seconds versus 1.25 seconds to first audio is an audible difference in a phone conversation. Customer service applications tend to be latency-sensitive, because a pause of even half a second registers as hesitation to callers.

The upgrade is automatic; no code changes are required. Any developer already calling that endpoint gets version 2.0 behavior automatically. Those who need to hold on version 1.0, for regression testing or established prompt calibration, pin the explicit model string.

xAI controls both the model and the deployment, so the conversion numbers are not third-party verified. The value is that a live telephony deployment exists at all, which is stronger evidence than a lab benchmark alone.

What to watch next

Two things will give the benchmark a more durable read. Independent developers publishing latency data from live traffic will clarify whether 0.70 seconds holds outside xAI's own infrastructure. Separately, Artificial Analysis has a public speech-to-speech leaderboard; a full published evaluation there, separate from xAI's own announcement, would give the 82.9% figure an independent endorsement it does not yet have.

Sources