xAI shipped a new voice model. The line that matters to you is dated August 5.

  • AI
  • Posted on1 month ago

xAI shipped a new voice model. The line that matters to you is dated August 5.

Bir çağrı merkezinde kulaklıklı mikrofon takmış, sıra halinde bilgisayar başında oturan çalışanlar

xAI released a new speech-to-speech model for voice agents. The numbers are good. But if you run a voice assistant in production, the line in that announcement that matters most to you is a calendar date.

%82,9
SPEECH-TO-SPEECH INDEX (1.0: 75.7%)
0.70s
TIME TO FIRST AUDIO (1.0: 1.25S)
0,4x
REASONING TOKEN USE (1.0 = 1.0X)
Aug 5
GROK-VOICE-LATEST MOVES TO 2.0

What xAI announced

On July 29, xAI announced Grok Voice Think Fast 2.0, its new speech-to-speech model aimed at developers building voice agents. It is priced at $0.08 per minute of audio.

On the speech-to-speech index published by the independent evaluator Artificial Analysis, 2.0 took the top spot with 82.9%. The previous version scored 75.7%, GPT-Realtime-2.1 scored 79.1%, and Gemini 3.1 Flash scored 69.5%.

It helps to know how that index is built, because a single number hides three separate abilities. The Artificial Analysis index is made of three equally weighted parts: speech reasoning (Big Bench Audio), conversational dynamics (a Full Duplex Bench subset), and agentic performance (τ-Voice). Conversational dynamics covers pause handling, turn taking, interruption and backchannel handling. τ-Voice looks at end-to-end customer service task completion across airline, retail and telecom scenarios.

Split that way, the table gets more interesting. On the agentic side 2.0 leads clearly with 56.5%, against 45.7% for GPT-Realtime-2.1 and 37.7% for Gemini 3.1 Flash. On conversational dynamics the picture changes. 2.0 scores 95.1% while GPT-Realtime-2.1 sits slightly ahead at 95.7%. Both land in the same band, so on that axis the two models are effectively level.

Latency improved in a way you can feel. Time to first audio dropped from 1.25 seconds to 0.70 seconds. On the same measure GPT-Realtime-2.1 comes in at 1.21 seconds and Gemini 3.1 Flash at 2.99 seconds. xAI also reports that median reasoning token use fell to 0.4 times that of 1.0, and that transcription accuracy tested across 24 languages improved by 1.4 times over 1.0.

The line on the calendar

Gri bir zemin üzerinde duran, beyaz kadranlı klasik analog kronometre

Inside the announcement there is a one-line calendar item: on August 5, 2026 the The grok-voice-latest alias moves from grok-voice-think-fast-1.0to grok-voice-think-fast-2.0‘a geçecek.

If that alias is what your code says, the model underneath your voice assistant changes that day without you doing anything. The date of the upgrade sits on your vendor's calendar.

Here is the subtlety: a better model is not automatically a model that behaves the same. On a live phone line three things shift at once.

  • The latency profile. When time to first audio halves, the rhythm of turn taking changes with it. Your barge-in thresholds, silence timeouts and "I have not finished speaking" handling were tuned against a different model.
  • Tool call timing. With reasoning token use down to 0.4 times, tool calls land earlier in the conversation. A lookup can return before the agent finishes its first sentence. That is good news, unless your flow assumed the answer would arrive later.
  • What transcription hands you. When transcription behaviour changes, so does the input your intent classifier sees. If you trained that classifier on the old output, or picked its thresholds against it, those thresholds are now looking at a different distribution.

None of this counts as a mistake by xAI. Moving is what an alias is for, and xAI published the switch date in advance. The point is what using that alias in production quietly means: you have handed the version decision to your vendor.

Same benchmark family, two different numbers

Something turned up during research that we did not expect, and it may be the most useful part of this piece.

When 1.0 launched in April, coverage based on xAI's own announcement put its τ-voice Bench score at 67.3%. MarkTechPost carried that figure in the headline of its April 25 article. Artificial Analysis measures the τ-Voice component of that same 1.0 release at 52.1%.

Same benchmark family, two evaluators, 15 points apart. We do not know where that gap comes from. It could be the evaluation harness, the subset, the runtime settings or something else entirely. The sources do not explain it, so we are not going to guess here.

The operator takeaway holds anyway. A score published by the vendor and a score measured by an independent lab can diverge substantially. Both are useful signal. Neither is your traffic.

How a team running voice assistants on phone lines reads this

At TAO AI LAB we build voice assistants that answer real phone lines for businesses, in Turkish and English. We run on Vapi, and we build the system with its Squads feature: a separate assistant per department.

That topology makes the model question bigger. Per Vapi's documentation the model is configured per assistant, so each department's assistant carries its own model setting. At the squad level there is membersOverrides which overrides the settings of every assistant in a squad at once without touching the underlying assistants. So the model can be defined in two layers. If you are going to pin a version you need to know both, because Vapi's API definition applies the per-assistant setting first and then lays the global override on top of it.

Then there are the handoffs. In Squads, assistants pass the call to each other through a handoff tool, and conversation context travels with the transfer, with you choosing how much history moves at each hop. A handoff is the moment in a call that is most sensitive to timing. When the latency profile shifts, that is most likely where you see the first thing break.

Two practices follow from this on our side.

First, pin explicit versions in production. An alias is convenient in development. On a live line the model name is a configuration value that can change behaviour, and configuration values get pinned.

Second, treat a model change as a deployment. Moving to a new version is work with a date, a verification run and a rollback plan. An upgrade that arrives on its own skips all three.

There is a third practice this picture points to: a fixed set of test calls. You take a set number of typical calls from your recordings, run them against the new model and compare the results with what the old one produced. A benchmark tells you what a model does on an average task. It says nothing about what it does with your customer's accent, your product names and the noise on your line. Only your own recordings tell you that.

My honest read, and I would like to hear where I am wrong. In voice agents the competition keeps moving away from model choice and towards the layer around the model. A 0.70 second first response is available to everyone. Turning it into a conversation that holds together is your job.

What this does not settle yet

Three limits keep this honest. First, we could not find the per-minute price of 1.0 in a verifiable source, so we are saying nothing about how pricing changed between versions. Second, Artificial Analysis is a single independent evaluator and its index rests on three datasets; a useful reference, though no substitute for your own call traffic. Third, we could not reach xAI's own announcement page, so the figures here come from the independent evaluator's site and from coverage dated July 29.

The direction is clear enough. Voice models keep getting faster and cheaper, and release cycles keep getting shorter. The teams that stay upright at this pace will be the ones who know in advance what happens when the model changes, as much as the ones who try each new model first.

Frequently asked questions

What is Grok Voice Think Fast 2.0?
A speech-to-speech model xAI announced on July 29, 2026. It targets developers building voice agents and is priced at $0.08 per minute of audio.

What happens on August 5, 2026?
The grok-voice-latest alias moves from version 1.0 to version 2.0 on its own. Applications using that alias will connect to the new model after that date.

Does Think Fast 2.0 lead on every measure?
No. It leads on the overall index and on agentic task completion. On the conversational dynamics component it scores 95.1%, slightly behind GPT-Realtime-2.1 at 95.7%.

Should I pin a version instead of using the alias?
In production, a pinned version leaves the timing of an upgrade with you. An alias brings the new model automatically, and in exchange your vendor sets the switch date.

A voice agent is only as steady as the model underneath it. The layer that creates that steadiness is what we build. If you run voice or agent systems, one question: who sets the date of your next model upgrade, you or your vendor? Tell me in the comments.

Sources

Leave A Comment

Categories