xAI's Grok Voice Think Fast 1.0 Tops τ-Voice Bench, Powers Starlink

xAI launched Grok Voice Think Fast 1.0 last week, and the benchmark numbers are unusually decisive. The new flagship voice model scored 67.3% on τ-voice Bench — nearly twice the score of Google's Gemini 3.1 Flash Live (43.8%) and almost double OpenAI's GPT Realtime 1.5 (35.3%).

The model is already in production at scale, handling phone sales and customer support for Starlink, where xAI says it converts 20% of inbound sales calls and resolves 70% of support tickets without human handoff.

What "Think Fast" Actually Means

Voice agents have been getting better, but most still pause noticeably while they "think" — the latency between when you finish talking and when the model responds. That delay breaks the natural rhythm of conversation and tanks customer satisfaction scores.

Grok Voice Think Fast 1.0 runs reasoning in the background, in parallel with the conversation. The model can work through a multi-step query while still listening, interrupting, and turn-taking like a human would. xAI describes it as "full-duplex" — speech in and speech out happening simultaneously, not sequentially.

Why this matters: this is the difference between a voice agent that feels like a phone tree and one that feels like a human assistant. For high-volume call centers, that difference is the entire business case.

The τ-Voice Bench Score in Context

τ-voice Bench is the new standard benchmark for full-duplex voice agents. It evaluates models under realistic conditions: background noise, accented English, mid-sentence interruptions, and natural turn-taking. Grok Voice Think Fast's 67.3% score isn't a small lead — it's roughly 24 points ahead of Gemini and 32 points ahead of GPT Realtime.

For comparison, xAI's previous Grok Voice Fast 1.0 scored 38.3%. The Think Fast version represents a 29-point jump in less than a quarter, which is the kind of step-change that usually only happens with a new architecture.

The model also handles structured data extraction — collecting email addresses, phone numbers, account numbers, and street addresses — with high accuracy even when callers speak quickly or with strong accents. That's the capability that makes it usable for actual customer service work.

Starlink as the Reference Customer

Starlink is xAI's flagship deployment and the proof point xAI is leaning on. The integration uses 28 separate tools across hundreds of support and sales workflows, including billing lookups, account modifications, technical troubleshooting, and order processing.

The reported metrics: 20% sales conversion rate on inbound inquiries and 70% autonomous resolution of customer support tickets. Both are well above industry norms for AI voice agents, which typically convert 5-10% of sales calls and resolve 30-40% of support tickets autonomously.

Starlink's scale matters here. The service has millions of subscribers across more than 100 countries, generating call volumes that would normally require thousands of human agents. If Grok Voice can sustain those metrics across that volume, it's a serious commercial signal.

API Pricing and Availability

Grok Voice Think Fast 1.0 is available now through the xAI API. The company is positioning it for enterprise voice deployments — customer support, outbound sales, appointment scheduling, and structured data collection.

xAI hasn't published full pricing, but the model is part of the same enterprise voice tier that powers the Starlink deployment. Companies interested in piloting it can request access through xAI's enterprise sales channel, with early integrations focused on customer support providers and contact-center platforms.

The Competitive Picture

Voice agents have become one of the most contested fronts in enterprise AI. OpenAI ships GPT Realtime through ChatGPT Enterprise. Google offers Gemini Flash Live for developer integrations. ElevenLabs and Deepgram own the speech-to-speech infrastructure layer. Anthropic has not yet shipped a real-time voice product.

What xAI's launch does is set a new bar. A 67.3% τ-voice Bench score means OpenAI and Google now have a public number to beat — and the gap is large enough that catching up will likely require architectural changes, not incremental tuning.

For enterprise buyers, this is a useful inflection point. A year ago, voice agents were experimental. Today, they handle a growing share of inbound customer contact for major consumer services. The Starlink deployment shows what 'production' actually looks like at scale.

Industry Implications

The biggest near-term impact is in customer support. Companies running large contact centers — telecoms, airlines, retailers, financial services — now have a working reference implementation showing 70% autonomous resolution. That's a number that fundamentally changes the cost structure of customer support operations.

Outbound sales is the second obvious vertical. A 20% conversion rate on inbound inquiries is impressive, but the same underlying capabilities apply to outbound calling, appointment setting, and lead qualification.

The third vertical is healthcare and government services, where structured data collection over the phone is still largely manual. Grok Voice Think Fast's structured-data capabilities make those workflows newly automatable.

What's Next

Expect rapid follow-up launches from competitors. OpenAI's GPT Realtime is overdue for a major refresh, and Google has been hinting at a next-generation Gemini Flash Live model targeting full-duplex voice. Anthropic has not committed to a voice product timeline, but enterprise buyers are pressuring it to ship one.

For developers, the practical question is integration. Grok Voice Think Fast is API-first, with no native consumer app. Companies will need to build the front-end experience around it — phone routing, call recording, CRM integration — using their existing voice infrastructure providers.

The bottom line: voice agents just crossed a usability threshold for enterprise deployment. Grok Voice Think Fast 1.0 is the first model that genuinely sounds and feels like a competent human agent, and Starlink's metrics suggest it can deliver real ROI at scale. If you run a call center, this is the model to pilot first.