xAI
Grok Voice 3.0 tops most voice tests at the same $0.08 a minute
xAI released Grok Voice 3.0 in the Voice Agent API at the same per-minute price as Grok Voice 2.0, and it leads five of six benchmarks xAI cited.
xAI released Grok Voice 3.0, a speech-to-speech model that takes spoken audio in and answers in spoken audio, at the same price as the Grok Voice 2.0 model it replaces.
The model is available now in the Voice Agent API, xAI said in an announcement dated Oct. 14, 2026. The company called it “the world’s most intelligent speech-to-speech model.” It costs $0.08 per minute, or $4.80 per hour, the same as Grok Voice 2.0.
Agents that use the grok-voice-latest alias were moved to Grok Voice 3.0 automatically, xAI said, with no changes needed.
Benchmarks
Grok Voice 3.0 beat Grok Voice 2.0 on every test xAI listed, and beat OpenAI’s GPT-Live-1 and Google’s Gemini 3.8 Live on all but one. Most of the results come from Artificial Analysis and are dated Oct. 13, 2026. xAI’s page gives no other run dates.
| Test | Grok Voice 3.0 | Grok Voice 2.0 | GPT-Live-1 | Gemini 3.8 Live |
|---|---|---|---|---|
| Speech-to-Speech Index | 86.9 | 81.3 | 83.2 | 82.6 |
| Big Bench Audio | 98.5% | 97.2% | 90.1% | 97.7% |
| Full Duplex Bench | 97.0% | 95.1% | 94.9% | 91.9% |
| τ-Voice | 71.6% | 56.5% | 74.5% | 68.6% |
| Task Success Rate | 94.8% | 94.6% | 87.4% | 89.1% |
| Speech Agent Arena Elo | 1070 | 1011 | 1048 | 990 |
Big Bench Audio measures speech reasoning. Full Duplex Bench measures turn-taking and interruptions, meaning how well a model copes when both sides can talk at once. τ-Voice uses support calls with accents, noise and interruptions. GPT-Live-1 finished ahead on it.
Cost and Latency
Grok Voice 3.0 cost $0.025 per task on Big Bench Audio, according to Artificial Analysis. GPT-Live-1 cost $0.030 and GPT-Realtime-2.1 (High) cost $0.055.
xAI said the model has “class-leading latency at under one second.” Against GPT-Live-1 (Astra), xAI said, it reaches comparable τ-Voice accuracy at 40% lower latency. On ServiceNow’s EVA-Bench, xAI said Grok Voice 3.0 “achieves the highest accuracy score at 74%.” That result has not been independently confirmed.
Deployments
xAI said Grok Voice runs in Tesla vehicles, where it handles “millions of conversations a day,” and powers voice mode in Grok Bot.
Starlink call centers take about 20,000 customer calls a day, according to xAI. Roughly 70% are resolved with no human. About one in four sales calls ends in a purchase. One agent verifies customers, troubleshoots hardware, issues credits and ships replacement equipment. The deployment figures are company-reported and have not been independently verified.
Our View
For developers running voice agents, the alias matters most. The grok-voice-latest alias moved to Grok Voice 3.0 on its own, at an unchanged price. Developers got a new model with no code change and no new bill.
That makes the comparison their job. The τ-Voice test covers accents, background noise and interruptions, and xAI’s score on it rose from 56.5% to 71.6%. Developers whose callers fit that description have the most reason to look closely at what the new model does on their lines.
Replaying recorded calls is the way to look. Prompts and call flows written for Grok Voice 2.0 now run on a model that may answer differently. A recording shows where.
The upgrade pays off for a developer only when the replayed calls show better outcomes than Grok Voice 2.0 produced on the same recordings.