Voice is quickly becoming the most demanding test for AI in the enterprise. A chatbot can take a few seconds to think. A caller on the phone will not wait. On 14 October 2026, xAI released Grok Voice 3.0, the third generation of its speech-to-speech model, and it arrives with a clear claim: it is now the best-performing voice model on the market for building agents that handle real conversations and complete real tasks.
Elon Musk called it “the world’s leading conversational voice model”. That is a vendor talking. Yet independent benchmarking firm Artificial Analysis does rank it first, and xAI can point to a large production deployment inside its own group. Below we look at what has changed, how it compares to OpenAI and Google, and what it means for customer experience and contact centre leaders.
What Grok Voice 3.0 is, and who it is for
Grok Voice 3.0 is a speech-to-speech model. It takes audio in and produces audio out within a single model, instead of chaining speech recognition, a language model and text-to-speech. That architecture matters for business use. Each hand-off in a traditional pipeline adds latency and strips out signals such as tone, hesitation and interruptions, which are exactly the things that make a phone conversation feel natural or frustrating.
The model is aimed at developers and organisations building their own voice agents: customer service lines, sales and support desks, in-app assistants and in-car interfaces. It is available through xAI’s Voice Agent API. xAI did not announce changes to its consumer Grok app at launch. The company, which has since been merged into SpaceX, published the release through the SpaceXSI account on X.
Benchmarks: first overall, not first everywhere
On the Artificial Analysis Speech-to-Speech Index, dated 13 October 2026, Grok Voice 3.0 takes the top spot.
| Model | Speech-to-Speech Index |
|---|---|
| Grok Voice 3.0 (xAI) | 86.9 |
| GPT-Live-1 Astra, Medium (OpenAI) | 83.2 |
| Gemini 3.8 Live, Extended Thinking High (Google) | 82.6 |
| Grok Voice 2.0 (xAI) | 81.3 |
The index combines several components, and Grok Voice 3.0 improves on its predecessor across all of them:
- Spoken reasoning (Big Bench Audio): 98.5%, up from 97.2%.
- Conversational flow (Full Duplex Bench, covering pauses, turn-taking and interruptions): 97.0%, up from 95.1%.
- Task success in live calls: 94.8%, a marginal gain on 94.6%.
- Human preference (Speech Agent Arena Elo): 1070, up from 1011, though this places it third rather than first.
For customer service teams, the more telling numbers sit in task-oriented tests. On τ-Voice, which simulates customer support calls, Grok Voice 3.0 scores 71.6%, a big jump from the 56.5% of version 2.0. OpenAI’s GPT-Live-1 still leads that test at 74.5%, and xAI frames its own result as comparable performance at 40% lower latency.
ServiceNow’s EVA-Bench, built specifically for enterprise voice agents across customer service, IT service management and HR scenarios, puts Grok Voice 3.0 on top with 74% accuracy. Nitin Aggarwal, Senior Director of Product Management for Agentic Evals at ServiceNow, said the model “sets a new Pareto frontier on EVA-Bench”, meaning it currently offers the best trade-off between completing tasks correctly and delivering a good conversational experience.
Latency and cost
xAI promises responses in under a second. Artificial Analysis measured 0.81 seconds to first audio, well ahead of GPT-Live-1 at 1.34 seconds. It is, however, slightly slower than Grok Voice 2.0, which came in at 0.70 seconds. The new model spends a little more time reasoning before it speaks. For most service scenarios that is a sensible trade, since a correct answer at 0.8 seconds beats a wrong one at 0.7.
Pricing is unchanged at $0.08 per minute, or $4.80 per hour of conversation. On the Big Bench Audio run, Artificial Analysis put the average cost per task at $0.025, against $0.030 for GPT-Live-1 and $0.055 for GPT-Realtime-2.1. Tool usage is billed separately, for instance $5 per 1,000 web search calls and $2.50 per 1,000 file search calls.
Teams already on the grok-voice-latest alias were migrated automatically, and Grok Voice 2.0 is now deprecated. For production systems, pinning the explicit grok-voice-3.0 model name is the safer choice, so a future release does not silently change how your agent behaves.
Capabilities that matter for enterprise deployments
- Tool calling during the conversation: web and X search, search across your own documents, MCP servers and custom functions, such as looking up an order or rescheduling an appointment.
- Telephony via SIP with your existing phone numbers, keypad (DTMF) input and call transfer to a human agent.
- Multilingual support: the documentation lists 20 languages with native accents, while the product page claims more than 25. The model detects the caller’s language automatically. Other languages work “with varying degrees of accuracy”, so testing in your own markets is essential.
- Session controls for reasoning effort, voice choice, pronunciation of brand names and specialist terms, idle timeouts and session resumption.
- An EU region. Besides two US regions, the model runs in eu-west-1. For European organisations handling customer data, that simplifies data residency and GDPR discussions, though it does not replace a proper legal and privacy assessment.
- Scaling tiers: concurrent sessions start at 10 and rise to 200 based on cumulative API spend, with higher limits available on request.
Starlink and Tesla as the reference cases
xAI’s strongest argument comes from within the Musk group. According to the company, Grok Voice already handles millions of conversations a day across Tesla and SpaceX.
At Starlink, the model handles around 20,000 customer calls daily in multiple languages. About 70% are resolved without any human involvement. A single agent verifies the customer, troubleshoots hardware, issues account credits and ships replacement equipment. On sales calls, roughly one in four ends in a purchase.
In Tesla vehicles, drivers use Grok Voice hands-free to check calendars and reply to email through Connectors, and hand off larger tasks such as ordering coffee or booking a restaurant to Grok Bot.
These are impressive figures, but they are self-reported by xAI and SpaceX. No independent audit of the Starlink resolution rate or conversion figures has been published.
What this means for customer experience leaders
Within a few months, OpenAI launched GPT-Live-1, Google shipped Gemini 3.8 Live and xAI has now released Grok Voice 3.0. Each leads somewhere. For organisations evaluating voice agents, a few conclusions stand out.
Voice quality is no longer the differentiator. All leading models sound natural and respond within about a second. The real gap lies in reliability: does the agent complete the task, retrieve the right data and know when to escalate to a human? Task-based benchmarks such as τ-Voice and EVA-Bench say more about that than an overall leaderboard.
Integration is where the value is created. A model knows nothing about your customers, products or policies. The return on investment depends on how well it connects to your CRM, knowledge base, ticketing and order systems, and on the guardrails you put around it. That work typically takes longer than choosing a model.
Test in your own languages and markets. Benchmarks are largely run in English. In Europe, where a single contact centre may serve customers in Dutch, French, German and Italian, performance per language and per accent should be measured before anything goes live.
Plan for change. With three frontier voice models released in quick succession, lock-in is a real risk. An architecture that lets you swap the underlying model, combined with your own evaluation set of real customer scenarios, keeps your options open.
A strong contender in a crowded field
Grok Voice 3.0 is a clear step forward. It leads the most widely cited independent ranking, costs less per task than its main rivals and is already proven at scale in Starlink’s contact centre. It is not the best on every measure, it is a fraction slower than its predecessor and its most striking real-world numbers come from inside its own group. For any organisation building voice agents for customer service, it belongs on the shortlist. Whether it is the right choice depends on your languages, your systems and how much of the conversation you are ready to automate.
