Grok Voice 3.0: xAI's New AI Model Can Run Your Call Center

xAI just shipped an AI model that can hold a phone conversation without anyone noticing it's a machine — and it's already fielding tens of thousands of calls a day for one of Elon Musk's other companies.

The model is called Grok Voice 3.0, and it went live on October 14 through xAI's Voice Agent API. Unlike the usual setup where one system converts your speech to text, a separate language model figures out a reply, and a third tool converts that text back into speech, Grok Voice 3.0 does the whole thing in one pass — it listens to audio and replies with audio directly, without the detour through text in between. xAI says that's what lets it respond in under a second, handle someone interrupting mid-sentence, and manage natural back-and-forth pauses the way a human would.

Built for the phone, not the chatbot window

This isn't a consumer voice assistant like Siri or Alexa. xAI is pitching it squarely at businesses that want to automate phone support, contact centers, and in-app voice assistants. Developers connect to it through an API that's deliberately shaped like OpenAI's existing Realtime API, so a company already using that can reportedly switch over by changing a URL and an API key rather than rebuilding their voice stack from scratch.

The feature list reads like a checklist for anyone running a support line:

• Custom voices instead of one default robotic tone
• Live function calls, so the AI can actually look up an order or update a record mid-call, not just talk
• Built-in web search and support for MCP tools (a standard that lets an AI model plug into external software and data sources)
• Pricing held at the previous model's rate of $0.08 per minute

xAI claims Grok Voice 3.0 tops Artificial Analysis's speech-to-speech leaderboard, and says it's already running Tesla's and Starlink's customer support lines, handling roughly 20,000 calls a day with about 70% resolved without a human stepping in. Those are xAI's own numbers — no independent lab has verified them yet, so treat the figures as a claim from the company that built the thing, not an audited result.

A model that listens and replies in under a second, on a support line already fielding 20,000 calls a day — that's the pitch xAI is making to every company still running a human call queue.

Why this matters for India specifically

India runs the largest call center and business process outsourcing (BPO) industry in the world, employing well over a million people and generating tens of billions of dollars a year serving clients from the US, UK, and elsewhere. A voice AI model that's genuinely good enough to run first-line support at $0.08 a minute isn't an abstract AI milestone for that industry — it's a direct competitor to the entry-level phone support jobs that sector depends on. Indian voice AI players like Yellow.ai and Haptik have been building similar speech agents for domestic and global clients, but xAI's entry adds a well-funded, aggressively priced competitor to a market that was already consolidating.

There's a regulatory angle too. Indian banks and telecom companies already lean on voice bots for customer service, and regulators — the Telecom Regulatory Authority of India (TRAI) and the Reserve Bank of India (RBI) — have both pushed guidelines around disclosure and consent when an automated system, rather than a human, is handling a customer call. As speech-to-speech models get harder to distinguish from a real person, that disclosure requirement stops being a formality and starts being the only thing standing between a customer and genuine confusion about who, or what, they're talking to.

Code24 covered a related shift last month when Alibaba's Qwen-Audio 3.1 slashed voice AI API prices by 95% — between that and Grok Voice 3.0, the cost of running a voice AI agent has dropped sharply in just a few weeks, which only accelerates how fast companies can justify replacing human phone support.

What happens next

None of this means call center jobs vanish overnight — plenty of calls still need judgment, empathy, or the ability to escalate a genuinely messy problem, things these models aren't being sold as capable of yet. But the economics are shifting fast enough that any company running a large phone support operation, in India or anywhere else, now has a cheaper, faster automated option sitting right next to the human one. The real story here isn't the benchmark score. It's how quickly "good enough to answer the phone" became a commodity.