xAI released Grok Voice 3.0 on 14 October 2026, a speech-to-speech model offered through its Speech to Speech API. Agents using the grok-voice-latest alias were moved to it automatically, so existing integrations changed over with no work from developers.
The price stays at $0.08 a minute, or $4.80 an hour, the same as Grok Voice 2.0. TechEnclave adds that text input carries a separate $0.004 charge. On the Artificial Analysis Speech-to-Speech Index the model scores 86.9, against 81.3 for version 2. xAI also reports 71.6% on the τ-Voice benchmark, just under GPT-Live-1's 74.5%, and a 97% score on Full Duplex Bench.
Latency is where accounts differ. xAI says responses come in under a second, 40% lower than GPT-Live-1. Artificial Analysis measured 0.81 seconds, slower than the 0.70 seconds recorded for version 2.
xAI says Starlink's call centres take about 20,000 calls a day with the model, resolving roughly 70% without a human, and that the same model powers voice in Tesla vehicles and Grok Bot. xAI calls Grok Voice 3.0 the most intelligent speech-to-speech model and a record for resolving support calls. Its Big Bench Audio cost is $0.025 a task, against $0.030 for GPT-Live-1. TechEnclave calls the independent results a mixed trade-off rather than a gain on every measure, and says version 2.0 is now deprecated.
AI Weekly notes these figures are xAI's own, with no independent audit or published model card.
Sources · 3