Skip to content
NewsModel release

xAI Grok Voice Think Fast 2.0 tops speech-to-speech quality benchmark at 82.9%, passing GPT-Realtime-2.1

· by Pondero Newsdesk

The short version

xAI released Grok Voice Think Fast 2.0 on July 29, 2026, reaching 82.9% on Artificial Analysis's speech-to-speech quality index, ahead of GPT-Realtime-2.1 High at 79.1%. Time to first audio dropped to 0.70 seconds and the model is priced at $0.08 per audio minute.

xAI Grok Voice Think Fast 2.0 tops speech-to-speech quality benchmark at 82.9%, passing GPT-Realtime-2.1

xAI's Grok Voice Think Fast 2.0 scored 82.9% on Artificial Analysis's speech-to-speech quality index as of its July 29 release, 3.8 percentage points ahead of GPT-Realtime-2.1 High and 13.4 points ahead of Gemini 3.1 Flash High, per xAI's announcement. Time to first audio dropped from 1.25 seconds in the prior version to 0.70 seconds.

Benchmark results and what changed

Four benchmark categories show where Think Fast 2.0 moved relative to rivals and its predecessor, per xAI. On Artificial Analysis's Speech Reasoning Big Bench Audio, the new model reached 97.2%, a marginal step over Think Fast 1.0's 97.1%. On the agentic performance tau-voice Bench, Think Fast 2.0 scored 56.5% against GPT-Realtime-2.1 High's 45.7%, a 10.8-point advantage on tasks that require tool use across multiple turns.

Transcription accuracy shows the sharpest gains in difficult conditions. Word error rates fell 1.5-2x relative to Deepgram Nova 3 and ElevenLabs Scribe v2 across 24 languages in standard conditions. In noisy environments, the gap expands to roughly 10x, per xAI's internal evaluation. On reasoning efficiency, Think Fast 2.0 uses 0.4x the reasoning tokens of its predecessor. xAI says tool calls now typically complete before the agent finishes its first sentence.

Pricing sits at $0.08 per audio minute. The grok-voice-latest alias moves from Think Fast 1.0 to 2.0 on August 5, 2026. Developers running grok-voice-latest who need to stay on the prior version must pin grok-voice-think-fast-1.0 before that date.

Why it matters

xAI included A/B testing data from Starlink's customer support line to back the benchmark numbers with a real deployment. Per xAI, switching to Think Fast 2.0 on that line produced a significant increase in sales conversion rate and support containment rate. xAI did not publish the exact figures, and the test ran on xAI's own infrastructure. That makes the data a vendor case study, not an independent finding, and should be read accordingly.

For voice-agent operators, the two most actionable gains are the 0.70-second time to first audio and the roughly 10x WER reduction in noisy settings. Telephony and customer-support deployments that deal with background noise will see the biggest benefit from the accuracy improvement. The alias migration on August 5 is the immediate deadline: teams on grok-voice-latest have days, not weeks, to run a regression check before their prompts switch to a new model version.

Three voice API tiers now have a published benchmark position. Think Fast 2.0 leads the Artificial Analysis speech-to-speech index. GPT-Realtime-2.1 High sits in second at 79.1%. Gemini 3.1 Flash High trails at 69.5%. The quality ranking is clearer than it was before July 29, though pricing and latency tradeoffs still vary by deployment type.

What to watch next

Eleven Labs, Hume, and Cartesia have not published counter-benchmark results on the Artificial Analysis index. Scores from any of those providers in the next 30 days would contest Think Fast 2.0's top position. xAI has also signaled work on Grok 4 Aurora, described as a multimodal voice model, with no timeline or capability detail announced.

Sources