xAI Moved Every grok-voice-latest Agent to Grok Voice 3.0 on 14 October, at the Same $0.08 a Minute
- What happened
- xAI shipped Grok Voice 3.0 at $0.08 per minute on 14 October and switched every agent using grok-voice-latest to it automatically.
- Why it matters
- A voice agent on the alias changed models underneath with no action from its builder, and xAI's post gives no pinned id for either version.
- What to do
- Re-test any production voice agent on grok-voice-latest now: prompts, tool calls, latency and tone.
Our take: Grok Voice 3.0 arrived through an alias that moves on its own, which is fine for a demo and risky in production. If a live voice agent runs on grok-voice-latest, it changed models on 14 October whether you were ready or not. Re-test it this week. The price did not move, so the only cost is the testing.
Affected model: Grok Voice 3.0, which our directory rates Conditional: we have not tested it, and the alias moved to it without asking anyone.
What happened
xAI's post, dated 14 October 2026, says: "Grok Voice 3.0 is available today in the Speech to Speech API at $0.08 / min ($4.80 / hr), the same price as Grok Voice 2.0." It adds: "Agents using grok-voice-latest have already been switched over automatically, with no changes needed."
xAI quotes these Artificial Analysis figures:
| Metric | Grok Voice 3.0 | Grok Voice 2.0 |
|---|---|---|
| Speech-to-Speech Index | 86.9 | 81.3 |
| τ-Voice success rate | 71.6% | 56.5% |
| Cost per task (Big Bench Audio) | $0.025 | not given |
These are Artificial Analysis numbers as relayed by xAI. We did not read them on Artificial Analysis, and we have not run the model ourselves.
Why it matters
The upgrade was not opt-in. Any agent pointed at grok-voice-latest is now on a different model, and a 15-point jump on τ-Voice means behaviour changed, not just scores. Prompts tuned for 2.0, tool-call patterns, latency budgets and voice tone can all shift under a new model even when the headline numbers go up.
The launch post also names no model id for Grok Voice 3.0 beyond the grok-voice-latest alias, and none for 2.0. From the post alone, we cannot tell you how to pin the old version or lock to 3.0 so the next release does not move you again.
What changes for you
- Running a production voice agent on
grok-voice-latest: re-run your test calls now. Check prompts, tool calls, latency and tone against what you shipped. - Need stability: check xAI's API docs for a pinned model id before assuming you can stay on one version. The launch post gives none.
- Budget: unchanged at $0.08 per minute ($4.80 per hour).
- Choosing a voice model: the benchmark gap over 2.0 is large on paper, but it is one evaluator's numbers relayed by the vendor. Test on your own calls before you commit.
What to do
- 1 Re-run your test calls on every voice agent that uses grok-voice-latest; it is already on Grok Voice 3.0.
- 2 Check xAI's API docs for a pinned model id before relying on staying on one version; the launch post names none.
- 3 Budget is unchanged at $0.08 per minute ($4.80 per hour).
Affected tools & models
Never need to catch up again
The weekly delta: only verdict changes and act-now items. No digest filler.