Grok Voice 3.0 makes production voice agents worth retesting
The new speech-to-speech model leads an independent index at the same price, but its automatic alias upgrade is a production risk.
What shipped
xAI released Grok Voice 3.0 on October 14 through its Voice Agent API at $0.08 per minute, unchanged from the previous version. It scores 86.9 on Artificial Analysis’s Speech-to-Speech Index, up from 81.3 for Voice 2.0, with 0.81 seconds to first audio.
Where value sits
The useful gain is not simply better-sounding speech. Grok Voice 3.0 reaches 71.6% on τ-Voice, a benchmark for multi-step support calls, versus 56.5% for its predecessor. If that holds on real workflows, the same per-minute price could produce a lower cost per resolved request.
What we would do
xAI automatically moved grok-voice-latest to the new model and deprecated Voice 2.0. We would pin the exact version, replay production calls offline and measure completed tasks, escalations, latency and outcome cost before allowing any broad traffic shift.
Grok Voice 3.0 is not a reason to rebuild a contact centre around one vendor, but it is the first voice release in months that we would move into a controlled production bake-off immediately. At the same $0.08 per minute as Voice 2.0, the independent benchmark gain shifts the question from whether native speech-to-speech is affordable to whether it can reduce failed calls, incorrect tool actions and human handoffs enough to beat a mature STT–LLM–TTS stack on cost per resolved task. We would not trust the leaderboard alone, and we especially would not run grok-voice-latest in a system that must behave predictably: xAI moved that alias automatically, so production behaviour can change without our deploy. Virtual Arc would pin Grok Voice 3.0, replay a representative call set, measure tool-call correctness, escalation rate, p95 time to resolution and cost per successful outcome, then route only low-risk intents until the logs prove it deserves more. The model is worth touching now; an unpinned migration is not.