Grok Voice 3.0 Is Live, Voice 2.0 Deprecated: Migration Checklist
Grok Voice 3.0 is a speech-to-speech model xAI released on October 14, 2026. It is available in the Voice Agent API at the same price as 2.0. The most direct impact for developers: the grok-voice-latest alias now points to 3.0, so sessions that use the alias switched models without any code change. grok-voice-2.0 is marked deprecated, so projects that hard-code the version name need to schedule a migration.
Everything below is current as of October 2026 and follows the official xAI announcement and the official release notes.
What exactly changed?
What can be verified from the official announcement page and the release notes:
- The model is named
grok-voice-3.0. It launched on the Voice Agent API on October 14, priced at $0.08/minute ($4.80/hour), the same as 2.0. The API itself is unchanged. - xAI says latency is under one second (the announcement's wording is "under one second", with no measurement methodology given).
- The announcement cites the Artificial Analysis speech-to-speech index: 3.0 scores 86.9, 2.0 scores 81.3. This is a third-party benchmark, and it is the result xAI itself chose to cite. It does not tell you how the model will perform in your use case.
grok-voice-latestnow points togrok-voice-3.0, so sessions using the alias are upgraded automatically.grok-voice-2.0is deprecated. The official upgrade path is to changemodeltogrok-voice-3.0orgrok-voice-latest. The release notes give no specific shutdown date; check the official docs.
What are the two other related changes?
| Model | Change | Impact |
|---|---|---|
grok-voice-transcribe-1.0 | Reached end of life on October 2, 2026 | Requests are routed to grok-voice-transcribe-2.0 at the same price; xAI says accuracy is higher |
grok-imagine-image-quality | Retires on November 2, 2026 | Requests are handled by grok-imagine-image-2.0 with quality set to low; the price per image is lower, and the request and response formats are unchanged |
The transcription redirect is already in effect. If your transcripts show differences in wording or punctuation around early October, this may be the reason.
Alias or pinned version?
An alias is convenient, but the model can change under you without notice. A pinned version is stable, but you have to watch for deprecation announcements yourself.
- Production apps where conversation quality matters and you have regression tests: pin
grok-voice-3.0and upgrade on your own schedule. - Prototypes and internal tools:
grok-voice-latestis fine, as long as you accept that behavior will shift with the model. - Either way, log the actual model name used for each session, so you can line up a problem with the time it started.
Migration and regression test checklist
- Search your code, environment variables, and config for
grok-voice-2.0,grok-voice-transcribe-1.0, andgrok-imagine-image-quality. - Change
modeltogrok-voice-3.0, or make an explicit decision to stay on the alias. - Replay a batch of real calls or recordings. Focus on barge-in (does the model stop promptly when the user interrupts), turn-taking (does it talk over the user, or leave long silences), and whether tool calls still fire as expected.
- Compare how your prompts behave. The model changed, so prompt workarounds you wrote for 2.0's quirks may no longer be needed and could even backfire.
- If you budget by the minute, the price is unchanged, but call duration may shift if turn pacing changes. Spot-check your bill.
- Subscribe to the official changelog through /en/changes or RSS so the next deprecation doesn't slip past you. For methods, see How to find out right away when an LLM API price changes.
Real-time voice stuttering: is it the network?
Voice tolerates far less latency than text. xAI's under-one-second figure is a server-side number; what you actually feel also includes the network round trip between you and the API. If things feel "laggier" after the model change, first separate the model from the network path:
- Open /en/status/xai and check the endpoint probes and whether there have been recent anomalies. For how to read status pages, see How to read the OpenAI and Claude status pages; the approach is similar.
- Test first-connection and round-trip time on the cdnping home page. The first connection includes DNS and a TLS handshake; once a long-lived connection is up, round-trip time matters more.
- Check the latency signature at /en/ip. If the browser round trip and the TCP round trip from the exit differ a lot, there is a relay in the middle, which directly slows real-time audio.
- Break latency into connection, gateway round trip, and first token. See AI API latency breakdown; for an overall check, use the developer network self-check checklist.
If the status is normal, the path is clean, and the same recordings behave differently on 2.0 and 3.0 in your regression tests, then it is a difference in model behavior. Adjust your prompts using the checklist above.
Related reading on pricing
xAI prices its text models separately from voice; see Grok 4.7 API pricing and reasoning tiers. Voice and text models are billed separately, so don't mix them in one estimate.
