xAI released Grok Voice 3.0 on October 14, 2026, and it now ranks first on Artificial Analysis’s Speech-to-Speech Index. The model scores 86.9, ahead of OpenAI’s GPT-Live-1 and Google’s Gemini 3.8 Live, according to the announcement. The price stays at $0.08 a minute, the same as Grok Voice 2.0.
It is not first everywhere. On τ-Voice, the test of support calls with accents, noise and interruptions, GPT-Live-1 still scores higher. Here is what the numbers say, what it costs next to the other two live voice models, and what xAI says it already runs on.
On this page · 6 sectionsOpen
xAI released Grok Voice 3.0 on October 14, 2026, a speech-to-speech model for voice agents, in its Voice Agent API.
It scores 86.9 on Artificial Analysis’s Speech-to-Speech Index, first ahead of GPT-Live-1 (83.2) and Gemini 3.8 Live (82.6).
On τ-Voice, the support-call test, it scores 71.6% against GPT-Live-1’s 74.5%; xAI’s pitch is similar accuracy at 40% lower latency.
The price stays $0.08 a minute ($4.80 an hour), the same as Grok Voice 2.0; agents on grok-voice-latest moved over automatically.
xAI says Starlink’s call centers take 20,000 calls a day on Grok Voice, about 70% resolved with no human in the loop.
§ 01What xAI announced
Grok Voice 3.0 is a speech-to-speech model: it listens and answers in audio, for voice agents that take calls or talk inside apps. It is in xAI’s Voice Agent API today. “Agents using grok-voice-latest have already been switched over automatically, with no changes needed.” xAI’s model list now marks Grok Voice 2.0 as deprecated.
§ 02The benchmarks
xAI published Artificial Analysis’s comparison, dated October 13, 2026. When we checked Artificial Analysis’s public speech-to-speech page on October 14, it did not yet list Grok Voice 3.0, so the figures below are xAI’s citation of the firm’s testing.
| Benchmark | Grok Voice 3.0 | Grok Voice 2.0 | GPT-Live-1 (Astra, Medium) | Gemini 3.8 Live (Ext. Thinking, High) |
|---|---|---|---|---|
| Speech-to-Speech Index | 86.9 | 81.3 | 83.2 | 82.6 |
| Big Bench Audio (speech reasoning) | 98.5% | 97.2% | 90.1% | 97.7% |
| Speech Agent Arena Elo | 1070 | 1011 | 1048 | 990 |
| Task Success Rate | 94.8% | 94.6% | 87.4% | 89.1% |
| Full Duplex Bench | 97.0% | 95.1% | 94.9% | 91.9% |
| τ-Voice (support calls) | 71.6% | 56.5% | 74.5% | 68.6% |
Three things stand out. The jump from Grok Voice 2.0 is real: 5.6 points on the index and 15.1 points on τ-Voice. Grok leads five of the six rows, including the human-preference arena and Full Duplex Bench, which measures how a model handles pauses, turn-taking and interruptions. And the one row it loses is the support-call test, where GPT-Live-1 scores 74.5% to its 71.6%. xAI’s answer is speed: it says Grok gets there with 40% lower latency, under one second to first audio.
On ServiceNow’s EVA-Bench, 200+ enterprise voice scenarios scored on accuracy and experience, xAI reports 74% accuracy, the highest it lists. ServiceNow’s Nitin Aggarwal is quoted: “It pairs top accuracy with a natural, responsive conversation.”
§ 03What it costs
| Model | List price | What it covers |
|---|---|---|
| Grok Voice 3.0 | $0.08 a minute ($4.80 an hour), $0.004 per text input | The full speech-to-speech model |
| GPT-Live-1 | $0.05 a minute, billed per second | The voice only; your backend model bills on top |
| Gemini 3.8 Live | $0.005 a minute audio in, $0.018 a minute audio out | Audio priced by direction; thinking included in output |
Sources: xAI models and pricing, OpenAI GPT-Live-1 model page, Gemini API pricing.
Per minute, Grok has the highest list price of the three: 1.6 times GPT-Live-1’s. But the units differ. GPT-Live-1 hands its reasoning to a backend model you pay for separately, and Gemini bills each direction of audio on its own. Artificial Analysis measured cost per completed task on its Big Bench Audio run: Grok Voice 3.0 at $0.025, GPT-Live-1 (Astra, Medium) at $0.030 and GPT-Realtime-2.1 (High) at $0.055. On that measure Grok is about 17% cheaper than GPT-Live-1. For a call center, cost per resolved call is the number that matters, and it depends on your own call length and backend.
§ 04Where it already runs
xAI says Grok Voice 3.0 handles millions of conversations a day in Tesla vehicles, and powers Grok Bot’s voice mode. The most concrete claim is Starlink: its call centers answer 20,000 customer calls a day on Grok Voice, in multiple languages. Roughly 70% are resolved with no human in the loop, and one in four sales calls ends in a purchase before the caller hangs up, xAI says. One agent verifies customers, troubleshoots hardware, issues credits and ships replacement equipment. These are xAI’s own figures about its sister companies; no third party has published them.
§ 05What we are watching
- Whether Artificial Analysis’s public leaderboard shows the same order once other labs submit new models.
- Whether OpenAI or Google answer on price; Grok’s per-minute list price is the highest of the three.
- Independent results on τ-Voice and EVA-Bench, beyond the figures xAI published.
§ 06Sources
- xAI: Introducing Grok Voice 3.0 (October 14, 2026)
- xAI models and pricing
- OpenAI GPT-Live-1 model page and Gemini API pricing
- Artificial Analysis speech-to-speech leaderboard
- Our record of GPT-Live-1 and Gemini 3.8 Live prices
Q1How does Grok Voice 3.0 compare with GPT-Live-1?
On Artificial Analysis’s numbers, Grok Voice 3.0 leads the overall index (86.9 vs 83.2), Big Bench Audio, the Speech Agent Arena and Full Duplex Bench. GPT-Live-1 (Astra, Medium) leads τ-Voice, 74.5% to 71.6%. xAI says Grok reaches its τ-Voice score at 40% lower latency.
Q2Is $0.08 a minute expensive?
It is the highest list price per minute of the three. GPT-Live-1 is $0.05 a minute for the voice alone, with your backend model billed on top, and Gemini 3.8 Live bills $0.005 a minute of audio in and $0.018 out. On Artificial Analysis’s per-task cost, Grok came in lowest: $0.025 against $0.030 for GPT-Live-1.
Q3What changed from Grok Voice 2.0?
Every published score rose: the index from 81.3 to 86.9, τ-Voice from 56.5% to 71.6%, Full Duplex Bench from 95.1% to 97.0%. The price did not change, and xAI’s docs now mark Grok Voice 2.0 as deprecated.
Q4Who runs these benchmarks?
The comparison table comes from Artificial Analysis, dated October 13, 2026, as published in xAI’s announcement. EVA-Bench is ServiceNow’s open-source benchmark; xAI reports 74% accuracy there. We have not run our own tests.
Q5Does CellCog use Grok Voice?
No. CellCog’s AI employees take meetings and phone calls with their own live voice, and CellCog does not route to Grok Voice.