The public τ-voice Bench leaderboard has seen Gemini 3.1 Flash Live and GPT Realtime trading places for the past year. On April 25, xAI dropped a new model onto it, scoring 67.3%.
The runner-up, Gemini 3.1 Flash Live: 43.8%.
A gap of 23.5 percentage points isn't called a "lead" in LLM benchmarks; it's called "a different league."
What's more noteworthy than the score is that the model is already in production on Starlink's customer service system.
grok-voice-think-fast-1.0: The numbers first
The model is called grok-voice-think-fast-1.0, officially positioned as a "full-duplex voice agent." Full-duplex means it doesn't need to wait for you to finish speaking before replying; it can listen and talk simultaneously, like a normal human conversation.
τ-voice Bench scores by vertical:
| Domain | grok-voice-think-fast-1.0 | Runner-up |
|---|---|---|
| Retail | 62.3% | 45.6% |
| Aviation | 66.0% | 64.0% |
| Telecom | 73.7% | 40.4% |
The gap in telecom is the most staggering: 73.7% vs. 40.4%, nearly double. This isn't from stacking parameters; it's the result of xAI's extensive training specifically on customer service scenarios like electronics and account management.
It supports 25+ languages and remains stable in noisy environments, with accents, and when interrupted.
How it "thinks" without stuttering
The most interesting technical feature is called background reasoning.
Typical voice agents exhibit one of two behaviors when faced with a complex request: either they give a shallow answer (you can feel they're glossing over it) or they pause for a few seconds to think (you can feel the lag). Neither feels like talking to a person.
grok-voice-think-fast-1.0's approach is to perform reasoning in the background in parallel while responding to you. On the surface, a low-latency conversational model maintains fluency—"Uh-huh," "Right," "I see," "Let me check"—natural conversational fillers. Behind the scenes, another reasoning pass solves the actual problem you posed. When the reasoning pass produces a result, it's immediately integrated into the main dialogue stream.
xAI's official statement:
"Think in the background—handling complex queries and workflows with zero impact on response latency."
It sounds like marketing speak, but the telecom scenario on τ-voice Bench is precisely designed to test this: a user recites a long string of account numbers, subscription details, and address changes, and the model cannot pause for more than 200 milliseconds. grok-voice-think-fast-1.0's 73.7% score in this scenario is empirical proof that this mechanism works.
Starlink handed over all its customer service: Real, hard numbers
Theoretical benchmarks are boring. What's compelling is Starlink's deployment data.
Starlink integrated this voice agent into real sales and customer service phone lines. After running for a period, xAI disclosed several KPIs:
Sales conversion rate of 20%. Out of every five people calling for inquiries, one directly pays for Starlink service during the call.
To put that in perspective: in traditional telemarketing, even with specialized training and scripts, an average call-to-sale conversion rate of 5-8% is considered good. 20% is the level of top-tier human salespeople, and maintaining it consistently is very difficult—people get tired, emotional, and have bad days.
AI doesn't.
70% of customer service requests are fully resolved autonomously, without human intervention.
This figure is even more striking. It means that for 10 customer calls, the agent handles 7 of them completely on its own, resolving the issue before the call ends—bill inquiries, plan changes, basic technical troubleshooting, address modifications, service suspension and reconnection—all end-to-end, without transferring to a human.
The remaining 30% are escalated to human agents, but by that point, the voice agent has already packaged the customer's issue, account context, and attempted solutions—so the human agent can handle it directly without asking, "What's your account number? What have you tried?"
A single agent simultaneously operates 28 different internal tools.
This is the most noteworthy point. The term "agent" has been overused in the past year, but few can actually orchestrate dozens of tool calls in a production environment. In this Starlink deployment, a single conversation flow might involve: account system, billing system, device status query, coverage map, logistics tracking, refund process, technical tickets, on-site service scheduling, customer relationship records... In a 5-10 minute call, the agent might make 15 tool calls.
Why Starlink got it first
It's no coincidence.
xAI and Starlink are both within the SpaceX ecosystem. Last month, Musk merged SpaceX and xAI to file for an IPO (a $1.75 trillion valuation, the largest ever). This kind of "use your own product first" internal synergy is much faster than selling to Salesforce or Microsoft. Starlink dared to connect sales-critical phone lines to a new model because if something goes wrong, Musk can send a text to fix it.
This internal loop is also a goldmine for model training. Starlink's daily customer service call volume is in the millions, all flowing back to xAI as RLHF training data. Ordinary voice AI companies would need years to sign clients and navigate compliance to get real-world data at this scale. xAI took a shortcut.
What this changes
First, the leadership in the voice AI market has been reshuffled.
Previously, the top players in the voice agent space were OpenAI Realtime, Google Gemini Flash Live, ElevenLabs, Vapi, Bland, and others. The gap on τ-voice Bench this time isn't 1-2 points; it's a chasm. In the short term, Gemini and OpenAI will need to go back and fill capability gaps.
Second, the decision-making weight for B2B buyers has shifted.
Previously, enterprises chose voice vendors based on two main criteria: latency and voice quality. Now a third dimension has emerged: "can it close the loop end-to-end?" Starlink's 70% autonomous resolution rate has made this standard clear—voice agents that can close the loop and those that can't should be in two different price tiers.
Third, the ceiling for B2C user experience has been raised.
20% sales conversion, 70% autonomous resolution, operating 28 tools, 25 languages, zero-latency thinking. A year ago, these metrics were considered "theoretical limits" in the voice AI world. Now they are "measured results running in production." Every company using voice agents for customer service now has to rethink their baseline.
As for what Anthropic and OpenAI will do next, we probably won't have to wait long. They both have voice capabilities, but neither has yet presented a deployment case at the scale of "taking over all customer service lines for a SpaceX subsidiary." The τ-voice Bench curve will be very interesting to watch with the next update.
Sources: Grok Voice Think Fast 1.0 (xAI official announcement), CocoLoop, xAI Launches grok-voice-think-fast-1.0: Topping τ-voice Bench at 67.3% (MarkTechPost), xAI launches Grok Voice Think Fast 1.0 for voice agents (Testing Catalog), xAI Unveils Grok Voice AI That Thinks In Real Time While Handling Customer Support At Scale (Metaverse Post)