xAI released Grok Voice 3.0 on October 14, 2026, automatically upgrading applications that select its latest voice model. For businesses using that setting in phone support or voice assistants, the underlying model has changed without requiring an application-code update. The release notes confirm the switch.
The immediate decision is whether to accept automatic upgrades or choose a specific release. Voice 2.0 is now deprecated, but the migration guidance reviewed on October 14 gives no shutdown date.
Find out which model your service selects
Ask your developer or voice-service provider for the exact model setting—not just whether the product uses Grok. The current model-selection documentation distinguishes these three names:
Model setting | Documented position | Suggested next step |
|---|---|---|
| Now selects Voice 3.0 automatically. | Review current conversations and decide whether automatic upgrades suit your service. |
| Deprecated predecessor. | Plan and evaluate a move to Voice 3.0. |
| Explicitly selects Voice 3.0. | Use when you want to choose the release deliberately. |
xAI recommends versioned model names for production stability. Selecting a release makes the model choice explicit; it does not promise indefinite availability or unchanging service behavior.
The published notices do not establish what happened to calls already connected during the switch. If a third-party provider manages the service, confirm its settings and upgrade policy rather than assuming every Grok-powered product changed together.
Check customer outcomes, not just the connection
For an integration pinned to 2.0, xAI documents changing the model selection to grok-voice-3.0. The endpoint, session parameters and events stay the same. That describes software compatibility, not proof that conversations will be identical.
Artificial Analysis’s published results show Voice 3.0 completing 71.6% of simulated customer-service tasks on tau-Voice, compared with 56.5% for Voice 2.0. In its separate Big Bench Audio timing test, the first audio arrived after 0.81 seconds for 3.0 versus 0.70 seconds for 2.0. These are benchmark-provider measurements, not RohitAI tests or a forecast for your calls.
The practical lesson is to assess task completion and waiting time separately. Before accepting the new model for your service, compare representative calls while initially keeping the prompt, voice, tools and audio settings constant.
Drawing on xAI’s prompting guidance, useful acceptance checks include:
Understanding the caller. Try background noise, interruptions and corrected names or numbers. Check that unclear speech leads to clarification, not an invented answer.
Following business rules. Check answers against approved prices and policies. If simplifying the prompt, preserve those facts and the rules for transferring callers to a person.
Finishing the task. Verify the actual booking, account change or requested human transfer—not only the assistant’s spoken claim that it happened.
The rate is unchanged; the bill depends on usage
The published rate card lists both Voice 2.0 and Voice 3.0 at US$0.08 per minute, or US$4.80 per hour, plus US$0.004 per chargeable text input. These are API rates, not a quote for a third-party phone service.
Under the billing rules, default hands-free sessions using server_vad—automatic detection of when someone speaks—are charged for session duration. Push-to-talk sessions are charged for audio sent and received. Counting only spoken minutes therefore understates the default session’s billable time.
The text charge is not a fee for every spoken exchange. It applies to client conversation.item.create events, except tool-result items (function_call_output) and audio items. response.create is not itself a chargeable text input; generated audio still uses the audio meter. The model reference explains the exceptions.
Illustrative calculation, not measured usage: assume 1,000 hands-free sessions, each lasting five billable minutes and containing three chargeable text events. Using the published rates:
Audio: 1,000 × 5 × US$0.08 = US$400.
Text: 1,000 × 3 × US$0.004 = US$12.
Subtotal: US$412.
One extra connected minute per session would add US$80 across those 1,000 sessions, before any other changes. The example excludes telephony, hosting, tools and other services, retries, taxes, discounts and human follow-up. It is not a savings forecast.
Track total charges alongside correctly resolved cases and human follow-up. An unchanged per-minute rate alone cannot tell you whether the service costs less to operate.
If your business also uses Grok for images, see RohitAI’s guide to the separate image-model migration. Its retirement schedule does not establish a deadline for Voice 2.0.
Methodology: AI-assisted reporting and analysis based on xAI documentation and Artificial Analysis results, rechecked on October 14, 2026. RohitAI did not run voice-model tests. The cost example is arithmetic under stated assumptions, not observed performance.