On October 14, 2026, xAI announced Grok Voice 3.0 as its new speech-to-speech model. Independent Artificial Analysis numbers in the post: #1 on the Speech-to-Speech Index at 86.9, under one second class-leading latency, and 97% on Full Duplex Bench. Voice Agent API pricing stays $0.08/min — same as 2.0. Agents on grok-voice-latest already moved over.
Where it is already working
xAI says Grok Voice handles millions of Tesla conversations a day and powers Starlink call-center volume on the order of 20,000 calls/day, with roughly 70% resolved without a human. SiliconAngle covered the SpaceXSI framing the same day. EVA-Bench (ServiceNow) is cited at 74% accuracy, best when jointly scored with customer experience.
For builders, the useful bit is parity pricing with a real jump on turn-taking and task completion versus 2.0 (Index 81.3 → 86.9). τ-Voice at 71.6% with lower latency than the GPT-Live comparison in their chart is the “agent that can wait and interrupt sanely” story.
Practical note
If you already bill voice agents against Grok, check the dashboard after the auto-switch and re-run your accent/noise evals. Same price does not mean identical failure modes. I care about duplex behavior under barge-in more than leaderboard screenshots — but shipping #1 at the old rate is still the right kind of boring upgrade.
Voice agents fail in accents, overlapping speech, and tool latency — not in demo scripts. Re-test your hardest support call with 3.0 before you change prompts. Leaderboards are a starting line; production transcripts are the finish.
Have a blessed day!