xai has released grok voice 3.0, an upgraded speech-to-speech model for voice agents with built-in voices, voice cloning and real-time tool calling. it tops artificial analysis’s speech-to-speech index with a score of 86.9, while pricing remains unchanged at $0.08 per minute.
what's new in grok voice 3.0
the update brings five main changes:
- 28 built-in voices and voice cloning: developers can clone a voice from an audio clip of up to 120 seconds.
- tool calling: agents can use tools such as web search during conversations.
- realtime api: supports websocket connections, server-side voice activity detection (vad) and ephemeral tokens for client apps.
- integrations: includes examples for webrtc and twilio, with support for livekit, pipecat and voximplant.
- automatic updates: agents using the
grok-voice-latestalias switched to the new version automatically.
xai says grok voice already handles millions of conversations a day in tesla vehicles and starlink call centers. starlink alone takes 20,000 customer calls a day, and about 70% are resolved without a human.
grok voice 3.0 benchmark results
grok voice 3.0 scores 86.9 on artificial analysis's speech-to-speech index, compared with 81.3 for v2.0. it also ranks ahead of gpt-live-1 (83.2) and gemini 3.8 live (82.6). its τ-voice score rose from 56.5% to 71.6%, and its arena elo rating increased from 1011 to 1070.
on full duplex bench, which tests how well a model handles pauses, interruptions and turn-taking, v3.0 scored 97%, up from 95.1%. xai also reports sub-second latency and says the model is 40% faster than gpt-live-1 on τ-voice while achieving a comparable score.
improvements on other tests are smaller. big bench audio rose from 97.2% to 98.5%, while task success edged up from 94.6% to 94.8%. gpt-live-1 still leads on τ-voice, scoring 74.5%.
grok voice 3.0 is also cheaper to run. on artificial analysis's big bench audio run, it costs $0.025 per task, compared with $0.030 for gpt-live-1 and $0.055 for gpt-realtime-2.1.
the results point to clear gains in overall speech-to-speech performance, though v3.0 does not lead in every category.
sources
- introducing grok voice 3.0, xai – oct 14, 2026; artificial analysis benchmarks, τ-voice, full duplex bench, pricing
- voice api overview, xai docs – realtime endpoint, built-in voices, custom voices, ephemeral tokens, demo apps
- speech-to-speech leaderboard, artificial analysis – speech-to-speech index, arena elo, task success, cost per task
- τ-voice paper, sierra research on arxiv – benchmark for full-duplex voice agents on customer service tasks
- full-duplex-bench paper, arxiv – pause handling, turn-taking and interruptions
- big bench audio dataset, artificial analysis on hugging face – 1,000 spoken reasoning questions
