xAI's new voice model lands at the same price as the last one, takes the top spot on an independent speech index, and quietly upgrades anyone already pointed at its latest alias. Here's what Grok Voice 3.0 changes for voice agents.
xAI Launches Grok Voice 3.0 for Voice Agents
xAI released Grok Voice 3.0 on October 14, the newest version of the speech-to-speech model behind its Voice Agent API. Speech-to-speech means the model listens and replies in audio directly, without routing your words through a separate text step first, which is what makes voice agents feel like actual conversation instead of a transcription relay.
The company is pitching it at businesses that field customer support and sales calls with automated voice agents. Grok Voice 3.0 is available now in the Voice Agent API under the model name grok-voice-3.0, and if your agent points at the grok-voice-latest alias, you've already been moved over with no changes on your end. xAI now lists the older Grok Voice 2.0 as deprecated.
One naming note worth flagging: xAI is now part of Elon Musk's SpaceX, which renamed the AI unit SpaceXAI and then SpaceXSI earlier this month. You'll see both the xAI and SpaceXSI names in coverage of this launch.
Grok Voice 3.0 Takes the Top Spot on Speech Index
The strongest data point comes from Artificial Analysis, an independent testing firm that ranks speech-to-speech models on reasoning, task success, and listener votes in blind head-to-head calls. It scored Grok Voice 3.0 at 86.9 on its Speech-to-Speech Index.
That put it above OpenAI's GPT-Live-1, which scored 83.2 with GPT-6 Astra handling the reasoning, and Google's Gemini 3.8 Live at 82.6. On Full Duplex Bench, which grades how naturally a model handles pauses and interruptions, the new Grok model hit 97%.
OpenAI kept a narrow lead on one test. On Tau-Voice, built around support calls with heavy accents and background noise, GPT-Live-1 scored 74.5% against Grok Voice 3.0's 71.6%. xAI counters that its model reached first audio about 40% faster on that same test, which matters in a live call.
Those results come from Artificial Analysis, not xAI's own lab, which makes them more useful than a vendor-run table. They're still one firm's methodology, so they tell you how these models rank under those specific conditions, not how they'll behave on your traffic.
What Grok Voice 3.0 Costs Through the API
xAI kept the price flat. Grok Voice 3.0 runs $0.08 per minute, or $4.80 per hour, the same rate as Grok Voice 2.0. There's also a $0.004 charge per text input.
On cost per completed task, Artificial Analysis pegged a Big Bench Audio run at about 2.5 cents on Grok Voice 3.0, against 3 cents on GPT-Live-1 and 5.5 cents on OpenAI's GPT-Realtime-2.1. For a business running thousands of support calls a day, that spread adds up fast.
Grok Voice 3.0
- $0.08 per minute ($4.80/hr), same as 2.0
- ~2.5 cents per Artificial Analysis Big Bench Audio task
- 86.9 on the Artificial Analysis Speech-to-Speech Index
OpenAI's voice models
- GPT-Live-1 scored 83.2 on the same index
- ~3 cents per task for GPT-Live-1, ~5.5 cents for GPT-Realtime-2.1
- GPT-Live-1 still leads on Tau-Voice (74.5% vs 71.6%)
Where Grok Voice 3.0 Is Already Running
xAI is running the model in its own products, which gives a sense of the scale. Tesla vehicles account for millions of Grok Voice conversations a day, with drivers using it to check calendars and answer email. Bigger tasks like booking dinner route through Grok Bot, whose voice mode also runs on the new model.
At Starlink, SpaceX's satellite internet business, Grok Voice answers about 20,000 customer calls a day across several languages. xAI says roughly 70% get resolved without a human stepping in, and that one in four sales calls ends in a purchase before the caller hangs up. Those are company figures and haven't been independently verified.
Outside Artificial Analysis, Grok Voice 3.0 also led on ServiceNow's open-source EVA-Bench, scoring 74% across more than 200 enterprise scenarios. A ServiceNow product director said the model pairs top accuracy with a natural conversation. ServiceNow built the benchmark, so read that result with the usual care.
Should You Migrate Your Voice Agent to Grok Voice 3.0
If you're already on xAI's voice API through the latest alias, the switch is done and the price didn't move. There's little downside to testing it.
For anyone weighing a move from OpenAI or Google, the case rests on two things: the independent index edge and the lower cost per task. Test both on your own call recordings, especially accented and noisy ones, since that's where Grok Voice 3.0 trailed.
Grok Voice 3.0 is live now in the Voice Agent API. Start with a small slice of real calls before switching a full production queue, and keep an eye on whether OpenAI and Google answer with their own voice model updates in the coming weeks.
