xAI released Grok Voice 3.0 through its hosted Voice Agent API on October 14. For a team choosing a voice agent for customer support, Artificial Analysis's current speech-to-speech table shows a 15.1 percentage-point gain on simulated support tasks over Grok Voice 2.0, a small change in its separate task-success measure, and a slower first audio response.

xAI says the grok-voice-latest alias now routes to 3.0 automatically. The current API guide also offers a versioned grok-voice-3.0 name for teams that need to pin the release. The WebSocket endpoint, session parameters and events remain the same as in 2.0, according to xAI. No application migration is needed for users of the alias, although a new model behind a moving alias is a reason to rerun an application's own call tests.

The big change

  • What changed: Grok Voice 3.0 replaced 2.0 behind the grok-voice-latest API alias. Artificial Analysis now reports a 71.6% score on its simulated customer-support benchmark, up from 56.5% for 2.0.
  • Why it matters: That measure tests whether an agent reaches the correct database state across scripted airline, retail and telecom problems. It is more relevant to a support workflow than a general audio reasoning score, although it is still a simulation.
  • What to watch: The same evaluator measured 0.81 seconds to first audio for 3.0, against 0.70 seconds for 2.0. Teams should check their own response time, task completion and billable session length before switching a pinned deployment.

The support score and the response delay

Artificial Analysis's methodology describes its τ-Voice test as a simulated customer calling an agent equipped with a policy document and domain-specific tools. Its airline, retail and telecom tasks have defined final database states, and the score averages three trials. The current table gives Grok Voice 3.0 71.6%, compared with 56.5% for Grok Voice 2.0: a 15.1 percentage-point rise under those test conditions. The evaluation uses a simulated customer, generated speech, transcription and a model grader. It does not measure the share of real callers whose problems a deployed service resolves.

The separate Speech Agent Arena task-success figure moves only from 94.6% for 2.0 to 94.8% for 3.0. That figure counts eligible conversations with the correct final task-completing tool calls; it uses a different test from τ-Voice. The overall Artificial Analysis index rises from 81.3 to 86.9. Its current version gives equal weight to audio reasoning, τ-Voice, blind caller preference and Arena task success. None of those figures alone establishes a gain on every customer-service task.

For speed, Artificial Analysis reports average time to first audio on its Big Bench Audio questions. Grok Voice 3.0 takes 0.81 seconds and 2.0 takes 0.70 seconds. The new model therefore took 0.11 seconds longer to start speaking in that test, even though xAI calls its latency “under one second” in the release announcement. The measure ends at the first audio output; it is not a full-call latency or resolution-time measure.

What the listed voice price covers

xAI lists both versions at $0.08 per minute, or $4.80 per hour, for speech-to-speech audio. Its model pricing page adds $0.004 for each billable conversation.item.create text event. With the default server_vad turn detection, the audio meter counts session duration. With push-to-talk, it counts audio sent and received. An always-open ten-minute session at the listed duration rate therefore has $0.80 in audio charges before any separately billed text events. That calculation is based on xAI's rate card, not a BIG CHANGE invoice or a tested call.

Developers need an xAI API key for a server-side WebSocket connection, or a short-lived token for a client application. The API guide documents wss://api.x.ai/v1/realtime, a model query parameter, a session.update event for instructions, voice and turn detection, and streamed audio events in reply. xAI lists a ten-session concurrent limit per team and a 120-minute maximum session. A support team would still have to supply its own policies, tools, customer data boundaries and escalation workflow; the published benchmark configuration does not establish how that specific implementation will perform.

xAI also reports that Starlink handles 20,000 calls a day with Grok Voice and resolves about 70% without a human. Those are company-reported operating figures; Artificial Analysis's synthetic tests do not verify them. BIG CHANGE reviewed the release, API and benchmark documentation. We did not use the API, call a support agent or reproduce the evaluation.

Sources & further reading