Grok Voice 3.0 is xAI's new speech-to-speech model for its Voice Agent API, and any agent on grok-voice-latest already runs on it without a code change. The price per minute is unchanged. Treat it as a silent model upgrade in production: retest turn-taking, tool calls and hand-offs to people before your callers find the differences.
What was released
Grok Voice 3.0 was announced by xAI on 14 October 2026 in its Grok Voice 3.0 announcement. It is the successor to Grok Voice 2.0 and powers the Voice Agent API, Grok Bot's voice mode, and Grok in Tesla vehicles. xAI says agents that use the grok-voice-latest model name have already been switched over, with no changes needed on the customer's side.
What actually changed
- Same price. The Voice Agent API keeps the same per-minute price as Grok Voice 2.0, according to xAI; check the xAI pricing page for the current rate.
- Better handling of real calls, by xAI's figures. The announcement reports gains over 2.0 on every benchmark it lists. The largest is on the τ-Voice support-call test, which uses varied accents, background noise and interruptions, where xAI reports 71.6 against 56.5 for Grok Voice 2.0. Smaller gains are reported for full-duplex conversation (pauses, turn-taking and interruptions) and for spoken reasoning.
- Response time. xAI describes latency as under one second and gives no per-turn breakdown.
- No new model name to adopt. The announcement names only
grok-voice-latest. It does not list a versioned name for 3.0 or say how long 2.0 stays available, so check the xAI voice API documentation before you plan a pinned rollout. - Production examples. xAI cites Starlink customer support as a user, with one agent verifying customers, troubleshooting hardware, issuing credits and sending replacement equipment.
All benchmark results above are xAI's own reporting; CodeDTX has not reproduced them.
What to retest this week
In CodeDTX's view, an automatic switch is still a model change, and voice agents are harder to watch than text agents because mistakes happen live, with a customer on the line. Work through these in order:
- Replay your evaluation calls. Run the recorded or scripted calls you used to approve the agent against the live model and compare outcomes, not just transcripts: did the caller's task finish, and did the agent call the right tools with the right values? Our guide to what AI agent evals catch covers how to set this up.
- Check turn-taking with your own callers. Better handling of interruptions changes when the agent stops talking and when it decides the caller has finished. Test with the accents, noise and phone lines your callers actually use, and listen for clipped confirmations of amounts, dates and account details.
- Recheck tool calls and side effects. If your agent can refund, cancel, book or change records, confirm that it still asks for confirmation before acting and that limits are enforced in your systems, not only in the prompt.
- Confirm escalation still triggers. A model that resolves more calls alone can also hold on to calls it should pass on. Check that complaints, vulnerable callers and low-confidence cases still reach a person, as described in routing AI agent decisions to human review.
- Watch production for a fortnight. Compare containment, transfer rate, call length and repeat calls before and after the switch, by call type, following how to monitor an AI agent in production.
When to hold back
If your voice agent works in a regulated area, such as payments, health or debt collection, an unannounced model change may need the same sign-off as a release you made yourself. Until you can pin a model version, record which model served each call, keep the previous prompt and settings ready to restore, and follow the steps in versioning an AI agent for rollback. If you were choosing a voice provider, the unchanged price makes Grok Voice 3.0 worth a trial, but judge it on your own call recordings rather than published benchmarks. And whatever the model, callers should be told they are speaking to an AI, as covered in when an AI agent must disclose it is AI.
Frequently asked questions
What is Grok Voice 3.0?
Grok Voice 3.0 is xAI's speech-to-speech model for voice agents, announced on 14 October 2026 as the successor to Grok Voice 2.0. It powers xAI's Voice Agent API, Grok Bot's voice mode and Grok in Tesla vehicles. xAI reports improvements over 2.0 in handling interruptions, background noise and support calls, with response times under one second and no change to the per-minute price.
Do I need to change my code to use Grok Voice 3.0?
Not if your agent uses the grok-voice-latest model name. xAI says those agents were switched to Grok Voice 3.0 automatically, with no changes needed. That also means your production agent changed model without a deploy on your side, so replay your evaluation calls, check tool calls and escalation paths, and compare call metrics before and after the switch date.
Does Grok Voice 3.0 cost more than Grok Voice 2.0?
No. xAI states that the Voice Agent API keeps the same per-minute price as Grok Voice 2.0. Your total bill can still change if call lengths change, for example if the agent resolves more calls itself or handles interruptions differently. Track average call length and transfer rate for a couple of weeks after the switch so any shift in spend has a clear cause.
Can I stay on Grok Voice 2.0?
The announcement does not name a versioned model for either release or say how long Grok Voice 2.0 remains available, so check the xAI voice API documentation and your console before relying on it. If you cannot pin a version, keep a record of which model handled each call and be ready to restore your previous prompts and settings if the retest shows problems.
