AI News5 min

Grok Voice 3.0: new speech-to-speech model, same $0.08 a minute, and an alias that upgraded you already

Grok Voice 3.0 is a new speech-to-speech model at the same $0.08 a minute, and the grok-voice-latest alias already moved your agent onto it. What changed, which numbers to trust, and a re-test checklist after a silent upgrade.

Erhan Timur27 October 2026Founder, Digital by Default

Share:X LinkedIn

Grok Voice 3.0: new speech-to-speech model, same $0.08 a minute, and an alias that upgraded you already

xAI released Grok Voice 3.0 on 14 October 2026. It is a speech-to-speech model for the Voice Agent API, priced at $0.08 per minute (about $4.80 an hour), the same as Voice 2.0. The API, endpoint and events are unchanged.

The line buyers should notice is this one: the grok-voice-latest alias now points to 3.0, and Voice 2.0 is deprecated. If your agent was built on the alias, it moved to the new model without a deploy. If you pinned grok-voice-2.0, you have a migration to schedule.

We covered the original API in Grok just shipped a Voice API and the no-code builder in Voice agents go no-code. This is the 3.0 note.

What changed

Speech-to-speech end to end. One model listens and speaks, rather than a chain of transcription, text model and synthesis. That is where the naturalness and the interruption handling come from
Tools during the call. Web search, X search and your own functions can be called mid-conversation. xAI's docs include guidance on avoiding audio overlap while a function runs, which tells you they have hit the same awkward silences everyone else has
Telephony. SIP support, so it sits behind carriers such as Twilio. The API also follows the OpenAI Realtime format, which makes it a near drop-in for teams already on that shape
Languages. Around 20 listed in the docs; the landing page says 25-plus. Other languages work with varying accuracy. Test yours
Where it already runs. xAI says it powers the Grok app's voice mode, voice inside Grok Bot, Tesla in-car voice and Starlink support calls

The numbers, with health warnings

xAI claims class-leading latency under one second. One independent review measured 0.81 seconds for 3.0 against 0.70 seconds for 2.0 on the same setup, and found 3.0 smarter on support calls but a beat slower
xAI cites a first place on Artificial Analysis's Speech-to-Speech Index at 86.9. A separate comparison in another report ranked an OpenAI configuration ahead on a different test. Both are vendor-adjacent and should be weighed as such

Our stance is the same as for every voice benchmark we have written about: run your own calls, with your own scripts, on your own phone lines.

Who should care

Teams already on the Grok Voice API via the alias. You are on 3.0. Re-run your regression calls this week, because behaviour changed under you
Teams pinned to 2.0. Plan the move; the deprecation is announced and the price is identical
Anyone building a phone agent. At $0.08 a minute for speech-to-speech with tools and SIP, Grok Voice is now priced the same as ElevenLabs's new Eleven Realtime models and sits in the same bracket as OpenAI's realtime line. The decision is about quality on your calls and the surrounding stack, not price
UK operators with accent-heavy or regional call volumes. Language coverage is decent; accent robustness is the thing to test

A re-test checklist after a silent upgrade

Latency on your trunk. Measure time-to-first-audio from a real PSTN call, not from a browser demo
Interruption handling. Talk over the agent mid-sentence and check it yields and recovers
Tool-call silence. Trigger your slowest function and listen to what the caller hears while it runs
Script adherence. Your compliance lines (recording notice, identity checks) must survive the model change
Accents and names. UK postcodes, surnames and street names are a known failure mode for every voice model
Cost per completed call. Tool calls bill separately; compare the invoice line, not the headline rate

Grok Voice 3.0 vs the alternatives

Option	Price (vendor-listed)	Notes
Grok Voice 3.0	$0.08/min	Speech-to-speech, tools, SIP, OpenAI-Realtime-shaped API
Eleven Realtime v1 Mini / Max (20 Oct)	$0.08/min	90-plus languages, 17,000-plus voices, OpenAI SDK compatible
OpenAI GPT-Realtime-2	Per-token	Broad ecosystem, Vapi and others integrate it
All-in-one platforms (Vapi, Bland, Synthflow)	Per-minute bundles	Telephony, orchestration and model in one bill

For the stack decision behind this table, read build or buy a phone agent in 2026.

Bottom line

Grok Voice 3.0 is a real model upgrade at no price change, which is the right kind of release. The operational story is the alias: a lot of agents moved to 3.0 without anyone deciding to. Pin a version if you need a stable voice, re-run your test calls either way, and keep the deprecation of 2.0 in your change calendar.