If your voice agent connects to xAI with model=grok-voice-latest and you tested it before October 14, it is not running the model you tested. On October 14, xAI launched Grok Voice 3.0, and its release notes say: "grok-voice-latest now points to grok-voice-3.0, so sessions that use the alias are upgraded automatically."

No deploy on your side. No config change. Your callers started talking to a different model once the alias moved.

That is the quiet half of the launch. The loud half is the scoreboard: xAI's new speech-to-speech model sits at the top of the Artificial Analysis Speech to Speech Index at 86.9, ahead of OpenAI's GPT-Live-1 (Astra, medium) at 83.2.

What changed on October 14

Here is the whole update, from xAI's own docs:

The endpoint stays the same WebSocket: the voice docs show wss://api.x.ai/v1/realtime?model=grok-voice-latest, with your API key in an Authorization: Bearer header. Their own example uses the alias, so any agent copied from that example is on it.

Editorial illustration of a railway switch lever flipping a single track from one platform to another while a small train of speech bubbles rolls through
An alias is a switch on the track. When the vendor throws it, every session on it changes direction.

The numbers, and who reported them

xAI's announcement publishes these numbers and credits them to Artificial Analysis, dated October 13, 2026:

Benchmark Grok Voice 3.0 Grok Voice 2.0 GPT-Live-1 (Astra, Medium)
Big Bench Audio 98.5% 97.2% 90.1%
Full Duplex Bench 97.0% 95.1% 94.9%
τ-Voice (agent success) 71.6% 56.5% 74.5%

Artificial Analysis also publishes an overall score on its own Speech to Speech leaderboard, which lists Grok Voice 3.0 first at 86.9, then GPT-Live-1 (Astra, medium) at 83.2, Gemini 3.8 Live Extended Thinking (High) at 82.6 and Grok Voice 2.0 at 81.3.

Look at the τ-Voice row, though. It's the agentic row, a benchmark of support calls with "diverse accents, background noise, and interruptions" per the announcement, and on the chart xAI published, GPT-Live-1 still leads it, 74.5% to 71.6%. xAI's pitch there is latency: the announcement claims 40% lower latency than GPT-Live-1 on that benchmark, and "class-leading latency at under one second." Unite.AI's launch coverage reports both scores and notes that xAI calls its result comparable.

The jump from 2.0 matters more for you than the lead over OpenAI. τ-Voice went from 56.5% to 71.6% between xAI's own two models (announcement). That's a large behavioral change. A model that completes more tasks also gets there by other routes, from when it yields the turn to when it hands off or asks again. Your prompts were tuned against the old routes.

I haven't run Grok Voice 3.0 myself. Treat every number above as reported, not measured by me.

Why an alias is a production risk

A -latest alias trades control for convenience. xAI's models page describes the general scheme: <modelname>-latest points to the latest version, while a dated <modelname>-<date> name "will not be updated and is for workflows that demand consistency." The same page recommends the aliases for most users. That's a reasonable default. It just means the alias promises the newest model and says nothing about keeping it the same.

The docs don't say whether a versioned ID like grok-voice-3.0 can change under you. Check the models page for a dated voice ID, and pin that if one is listed.

For a chat feature, a model swap shows up as slightly different wording. For a voice agent on a phone line, it can show up as:

  • a script that now interrupts callers sooner, or waits longer before answering
  • a different rate of escalations to humans
  • tool or web-search calls at different moments, with different costs
  • an instruction you relied on that the new model weighs differently

None of that is a claim about Grok Voice 3.0 in particular. It's what changes when any model changes under a prompt, and it's why you want to pick the moment it happens.

Editorial illustration of a silhouette holding a telephone handset whose coiled cord turns into a padlock, pinning a single version tag in place
Pin the version you tested. Move to the next one on your schedule, after your own checks.

Pin it, then upgrade on purpose

First, find out whether you're on the alias:

grep -rn "grok-voice-latest" .

If you are, pin the versioned ID. The URL shape comes from the voice docs; only the model value changes:

import os MODEL = "grok-voice-3.0" # pinned, not grok-voice-latest URL = f"wss://api.x.ai/v1/realtime?model={MODEL}" HEADERS = { "Authorization": f"Bearer {os.environ['XAI_API_KEY']}", }

If you were on the alias before October 14, you have been on 3.0 since that day. Pinning grok-voice-3.0 takes you off the alias, so the next time xAI moves it, your agent stays where it is. If you'd rather stay on what you originally tested, you can pin grok-voice-2.0, but it is deprecated per the release notes, so that only buys time.

Then run the upgrade like any other dependency bump:

  1. Keep a set of recorded or scripted test calls that cover your real flows: happy path, interruptions, a caller who changes their mind, a request that should escalate.
  2. Run them against the old pinned model and against grok-voice-3.0, with the same session.update config.
  3. Compare task completion, escalations, turn count and minutes per call. At $0.08 a minute on both versions (pricing), minutes per call is your cost signal.
  4. Switch the pinned ID when the new numbers are at least as good, and log the model ID on every session so you can tell the two apart afterwards.

Here is a trimmed version of the session config the voice docs show, which stays identical across both runs:

{ "type": "session.update", "session": { "voice": "eve", "instructions": "You are a support agent for Acme.", "turn_detection": { "type": "server_vad" } } }

The pattern is bigger than xAI

Every vendor with a moving alias can do this to you. One cheap guard: a CI check that fails when any -latest model ID shows up in production config. Then a model change only happens as a deploy you chose.

Do you pin model IDs in your voice or chat agents, or do you ride -latest and watch the dashboards?

If this was useful

A model swap that nobody deployed is exactly the kind of change you only catch if you trace your LLM calls and run evals against them. Observability for LLM Applications covers that: tracing, evals, cost and quality regressions for LLM features in production, so a silent upgrade shows up as a diff in your numbers instead of a surprise in your support queue.

And if you work with coding agents, I build Hermes IDE, a free, source-available IDE for running Claude Code, Codex and other agents side by side (GitHub).

Observability for LLM Applications — the book