Grok Voice 3.0 Tops the Speech Benchmarks at the Same Price
xAI's new speech-to-speech model leads Artificial Analysis's index, answers in under a second, and already handles about 20,000 Starlink support calls a day. Its price per minute did not move.
Voice is where the frontier labs are fighting hardest right now, and the fight has moved past demos. On October 14, 2026, xAI released Grok Voice 3.0, a speech-to-speech model it calls the most intelligent in the world. The benchmarks mostly support the claim. The more telling detail is what xAI says the model is already doing at work.
The numbers
xAI cites benchmark data from Artificial Analysis dated October 13. On its Speech-to-Speech Index, Grok Voice 3.0 scores 86.9, ranked first, ahead of OpenAI's GPT-Live-1 at 83.2 and Google's Gemini 3.8 Live at 82.6. Its predecessor, Grok Voice 2.0, scored 81.3.
Across the individual tests, as reported by xAI:
- Big Bench Audio (spoken reasoning): 98.5%, against 90.1% for GPT-Live-1 and 97.7% for Gemini 3.8 Live.
- Full Duplex Bench (turn-taking and interruptions): 97.0%, the top score.
- Task success rate: 94.8%, against 87.4% for GPT-Live-1.
- Speech Agent Arena (human preference Elo): 1070, against 1048 for GPT-Live-1.
- EVA-Bench from ServiceNow: 74% accuracy, the highest listed.
It does not win everything. On τ-Voice, a benchmark of support calls with accents and background noise, GPT-Live-1 scores 74.5% to Grok's 71.6%. xAI's own text calls the two comparable; its chart shows OpenAI slightly ahead. That is the benchmark closest to messy real-world calls, so it is the one to watch.
Speed and cost
Time to first audio is under one second. xAI also claims 40% lower latency than GPT-Live-1 on its τ-Voice comparison.
Pricing is where the release is most aggressive by omission: it did not change. Grok Voice 3.0 costs $0.08 per minute, or $4.80 an hour, the same as version 2.0. Developers using the grok-voice-latest endpoint were switched over automatically.
On a per-task basis, xAI puts the cost on Big Bench Audio at $0.025, against $0.030 for GPT-Live-1 and $0.055 for GPT-Realtime-2.1. Better scores at a lower cost per task is the combination that wins procurement decisions.
The deployments are the real pitch
Every lab can post a benchmark table. Few can point to production volume inside companies they effectively control. xAI can.
Starlink call centers use the model to handle about 20,000 customer calls a day across multiple languages. xAI says about 70% are resolved without a human, and one in four sales calls ends in a purchase. A single agent verifies customers, troubleshoots hardware, issues credits and ships replacement equipment.
Tesla vehicles use it for what xAI describes as millions of conversations a day, with drivers handing off simple tasks like calendar and email and larger ones like ordering coffee or booking dinner.
These are friendly customers within Elon Musk's orbit, so the numbers come with that caveat. But they are also large, live deployments that generate exactly the data a voice model needs to improve: real accents, real interruptions, real frustration.
The same model also powers voice mode in Grok Bot, xAI's consumer assistant, so the improvements land for ordinary users and API customers at the same time. That matters for the feedback loop. A model that serves consumers, car owners and a call center at once sees a far wider spread of speech than one tuned for a single channel, and the Starlink deployment in particular exposes it to frustrated customers with hardware problems, which is about as hard a test as conversational AI gets.
Where this leaves the voice race
The voice market has moved quickly this year. OpenAI shipped GPT-Live as a full-duplex model that listens and speaks at once. Google pushed extended thinking into Gemini Live and started a price war with aggressive per-hour rates. xAI's answer is to top the composite index and hold price flat.
The competitive dynamic now looks a lot like the text model market a year ago: frequent releases, narrow benchmark leads, and pricing pressure. The difference is that voice agents are being deployed directly into customer support, one of the largest pools of labor spending any AI product can target. A few points of task success translate into real headcount decisions.
What to watch
- Independent replication. The figures are xAI-reported on Artificial Analysis data. Third-party evaluations on noisy, accented, multi-turn calls will matter more than the composite index.
- The τ-Voice gap. If OpenAI keeps its lead on realistic support calls, enterprise buyers may weight that over the headline score.
- Languages. xAI says Starlink calls span multiple languages but has not published a list of supported languages.
The read
Grok Voice 3.0 is a strong release on paper, and the unchanged price makes it stronger. But the most important line in the announcement is not a benchmark. It is that a single voice agent already handles a full Starlink support workflow, from verification to shipping a replacement, and closes the ticket seven times out of ten. That is the number every call-center operator will be reading.