xAI Releases Grok Voice 3.0, Topping Speech-to-Speech Benchmarks at Unchanged Price

What happened

Per tech media reports on October 14–15, xAI, Elon Musk's AI company, released Grok Voice 3.0, a speech-to-speech model available through its Voice Agent API. Independent testing by Artificial Analysis dated October 13 placed the model first on its Speech-to-Speech Index with a score of 86.9.

Per xAI's announcement, Grok Voice 3.0 led GPT-Live-1 (Astra, Medium reasoning, 83.2), Gemini 3.8 Live (Extended Thinking, High, 82.6), and the previous Grok Voice 2.0 (81.3). The index is an equal-weighted combination of speech reasoning, τ-Voice agentic performance, blind head-to-head preference, and task success.

On components, speech reasoning (Big Bench Audio) reached 98.5% (vs 97.2% for 2.0) and Full Duplex Bench 97% (vs 95.1%); τ-Voice support-call task success rose from 56.5% for 2.0 to 71.6%, still below GPT-Live-1's 74.5%, but xAI says time to first audio is 40% lower (0.81s vs 1.34s). On ServiceNow's EVA-Bench for enterprise voice agents, accuracy was 74%.

Key facts

1. Price unchanged: the API stays at $0.08 per minute, same as 2.0; the grok-voice-latest alias now points to the new model, so existing users upgrade without code changes.
2. Live deployments: Grok Voice powers Grok Bot's voice mode and in-car conversations in Tesla vehicles; custom voices can be created from a reference clip of up to 120 seconds.
3. Benchmark caveat: Artificial Analysis's index combines four measures and is not a single score for every aspect of live customer-support calls; GPT-Live-1 still leads on τ-Voice alone.