Image: SpaceXAI
Grok Voice 3.0 is SpaceXAI's newest voice model, and the company is pitching it squarely at businesses that want AI to answer their phones. In its announcement, SpaceXAI calls it "the world's most intelligent speech-to-speech model" and says it replies in under a second while costing less than rival frontier models.
A speech-to-speech model listens and talks back directly, rather than turning your voice into text, thinking, and then reading an answer aloud. That matters for phone calls, where long pauses and clumsy interruptions quickly annoy callers.
Introducing Grok Voice 3.0, the world’s most intelligent speech-to-speech model.
— SpaceX Super Intelligence (@SpaceXSI) October 14, 2026
It excels at resolving customer calls, responds in under a second, and costs less than other frontier models. pic.twitter.com/cfDRrkQPEc
The model is aimed at developers, not regular Grok app users. It is available now through SpaceXAI's Voice Agent API at $0.08 a minute, or $4.80 an hour, which the company says is the same price as Grok Voice 2.0. Apps already pointed at the "grok-voice-latest" name were moved over automatically. SpaceXAI also says Grok Voice 3.0 powers voice mode in Grok Bot and the Grok assistant in Tesla cars.
The headline claims rest on testing by Artificial Analysis, an independent benchmarking firm, as presented by SpaceXAI. On that firm's combined Speech-to-Speech Index, Grok Voice 3.0 scores 86.9, ahead of OpenAI's GPT-Live-1 at 83.2 and Google's Gemini 3.8 Live at 82.6. It also leads on reasoning, task success and how naturally it handles pauses and interruptions.
Benchmark scores shared by SpaceXAI, citing Artificial Analysis. Image: SpaceXAI
It isn't a clean sweep, though. On τ-Voice, a test of how well agents handle support calls with accents, background noise and interruptions, SpaceXAI's own chart shows GPT-Live-1 ahead at 74.5% against Grok's 71.6%. SpaceXAI's argument is that Grok gets close while starting to speak about 40% faster.
On cost, SpaceXAI says an average task in Artificial Analysis's audio reasoning test cost about $0.025 with Grok Voice 3.0, versus $0.030 for GPT-Live-1 and $0.055 for OpenAI's GPT-Realtime-2.1.
Cost per task as shown by SpaceXAI. Image: SpaceXAI
The strongest real-world claim comes from inside Elon Musk's own companies. SpaceXAI says Starlink's call centres use Grok Voice for about 20,000 calls a day in several languages, that roughly 70% are resolved without a human, and that one in four sales calls ends in a purchase. ServiceNow, which runs its own voice-agent benchmark, is quoted saying the model scored highest on accuracy there. None of the Starlink figures have been independently checked, and they come from a sister company rather than an outside customer.
There are also practical limits that the announcement doesn't mention. SpaceXAI's developer documentation caps each team at 10 calls running at once by default, limits a single session to two hours, and lists only one US server region. Text messages sent during a call cost an extra $0.004 each. For a small business trying it out, that is fine. A company handling thousands of calls an hour would need to arrange higher limits, and callers far from the US may notice more delay than the "under a second" figure suggests.
The bigger picture is that voice is turning into a price fight. SpaceXAI has kept its per-minute rate flat while claiming better results, which puts pressure on OpenAI and Google. Businesses should still run their own test calls before trusting any leaderboard.
Comments