xAI Launches Grok Voice 3.0 and Deprecates Voice 2.0
The Voice Agent API keeps its listed price and request shape while latest-alias sessions switch automatically, leaving pinned 2.0 integrations with a migration decision.
Edited by Tyronne Panaino
xAI announced Grok Voice 3.0 on October 14 and made it available through the Voice Agent API. The company's release notes also deprecated the pinned `grok-voice-2.0` model, so teams building speech-to-speech agents now face two different paths: sessions using `grok-voice-latest` have already moved automatically, while integrations pinned to version 2.0 need an explicit model change.
The update matters most to developers and operators whose voice applications depend on stable model behavior. xAI says the API request shape and listed price remain unchanged, but an alias target change can still alter the model generating each response. That makes regression testing, latency monitoring and conversation-quality review practical release work even when no code migration is required.
Status
Confirmed, with medium internal confidence. Two official xAI-controlled pages confirm availability, alias routing and deprecation. This run did not independently test the API, fetch the underlying benchmark reports or verify xAI's production-usage claims.
What changed for API integrations
The xAI announcement lists Grok Voice 3.0 at $0.08 per minute, or $4.80 per hour, matching the stated price of Grok Voice 2.0. It says agents using `grok-voice-latest` were switched automatically and require no API change.
The xAI developer release notes make the migration boundary clearer. They identify `grok-voice-3.0` as the current speech-to-speech model behind the latest alias, describe the API as unchanged, and mark `grok-voice-2.0` as deprecated. A pinned integration can move either to the explicit 3.0 slug or to the rolling latest alias at the same listed price.
The fetched notes do not give Voice 2.0 an end-of-life date. Deprecation therefore creates a migration signal, not evidence that the old slug has already stopped working. Teams that need reproducible behavior can evaluate the versioned 3.0 slug before changing production traffic; teams that intentionally track the latest alias should recognize that the model transition has already happened.
The delta from Voice 2.0
xAI's July release notes had made Voice 2.0 available for speech-to-speech use and scheduled `grok-voice-latest` to route to it from August 5. The October update advances that same rolling alias to 3.0 and deprecates the previously pinned model. This is separate from Grok Voice Transcribe 2.0, xAI's speech-to-text model covered elsewhere: the new release generates live spoken responses rather than only turning audio into text.
For API buyers, the most concrete change is therefore lifecycle and routing behavior. The official pages support continuity in price and request shape, but they do not establish that every workload will preserve the same latency, conversational style, tool-use behavior or error profile after the alias switch. Those are application-level questions that require workload-specific testing.
What xAI claims about performance
xAI reports an 86.9 result on Artificial Analysis's Speech-to-Speech Index and 97% on Full Duplex Bench. It also says Grok Voice 3.0 reached 74% accuracy on ServiceNow's EVA-Bench. These figures are useful signposts, but this run fetched only xAI's summary, not the underlying benchmark pages, task sets or reproducibility material. They remain vendor-reported results rather than independently verified findings in this article.
xAI also says the model handles millions of conversations per day in Tesla vehicles and 20,000 customer calls per day in Starlink call centers. The announcement does not provide a measurement period or independent audit for those deployment figures. They show the scale xAI claims for its internal use, not a guarantee that an outside deployment will reproduce the same outcomes.
Practical implications and next checkpoint
Operators using the rolling alias should treat October 14 as the effective model-change date in their own telemetry. A useful comparison would separate request compatibility from response quality: confirm that existing calls still succeed, then review interruption handling, task completion, tool handoffs, latency and escalation behavior on the application's actual traffic pattern.
The next verifiable checkpoint is an xAI retirement date for `grok-voice-2.0`, followed by independent tests that disclose model settings, task mixes and measurement methods. Until then, the release is confirmed, while its broader performance advantage and production economics remain less certain than availability, price and alias routing.
Sources
Update note: Last reviewed 2026-10-15. We will revise this post if xAI publishes a Voice 2.0 retirement date or independent evidence materially changes the performance picture.
Sources
- xAI — Introducing Grok Voice 3.0 — official
- xAI Developer Docs — Release Notes — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.