Grok Voice 3.0 vs Grok Voice Think Fast 2.0: Free Upgrade, Real Gains, One Regression

The model Grok Voice 3.0 replaces is the same model you will find on Artificial Analysis under a different name. Search the board for Grok Voice Think Fast 2.0 and you will not find it; search for Grok Voice 2.0 and you will find 81.3, which is the figure the earlier model's own launch coverage reported for Think Fast 2.0 under the "High" configuration label. SpaceXAI's API calls the model ID grok-voice-2.0 and its release notes call the July generation Grok Voice 2.0, while the announcement that introduced it used the name Think Fast. So when SpaceXAI deprecated grok-voice-2.0 on 14 October 2026 and pointed grok-voice-latest at grok-voice-3.0, it retired the model its own leaderboard row and its own marketing name describe as two things.

The practical consequence is that an upgrade many teams have already received — silently, without a deployment, because the alias moved — is bigger than the naming confusion suggests and smaller than the index gap implies. Both models bill $0.08 per minute and $0.004 per text input, so the price did not change. Most of the quality did.

What the upgrade buys, component by component

Every figure in this section is Artificial Analysis's own testing, dated 13 October 2026 for Grok Voice 3.0 and from the same board for the outgoing generation.

• Speech to Speech Index — 86.9 for Grok Voice 3.0 vs 81.3 for Grok Voice Think Fast 2.0. A 5.6-point composite gain, which sounds like a version bump until you pull it apart.

• Agentic performance, τ-Voice — 71.6% vs 56.5%. This is where the release actually lives. Fifteen points of multi-turn task completion is the difference between a voice model that handles a scripted call and one that handles a call with branches.

• Speech Agent Arena preference — Elo 1070 vs 1011. Fifty-nine points of human preference, and the highest recorded on the board.

• Conversational dynamics, Full Duplex Bench — 97.0% vs 95.1%. Nearly two points on barge-in and turn-taking, which is the axis this model line was originally built to win.

• Speech reasoning, Big Bench Audio — 98.5% vs 97.2%. Both effectively at ceiling.

• Task success — 94.8% vs 94.6%. Flat within measurement. If your evaluation is a single-turn request completed correctly, these two models score the same and the upgrade is invisible.

• Time to first audio — 0.81 seconds for Grok Voice 3.0 vs 0.70 for Grok Voice Think Fast 2.0.

• Price — identical on both sides. $0.08 per minute, $4.80 per hour, plus $0.004 per text input.

The regression, and why it is the interesting part

Read the last two lines together. Grok Voice Think Fast 2.0 cut time to first audio from its own predecessor's 1.25 seconds to 0.70, and xAI sold that number hard — it was the fastest published figure in the category when it shipped in late July 2026, and it is still faster than anything else on the board. Grok Voice 3.0 gives back 0.11 seconds of it.

That is not a scandal and it is not an accident. It is the visible price of the τ-Voice jump. Whatever xAI added to the model to move multi-turn task completion from 56.5% to 71.6% costs time before the first audio frame, and no amount of prompt engineering recovers it, because the delay is inside the model rather than in front of it.

What makes it worth stating plainly is that it inverts the pitch. A team that chose Think Fast 2.0 did so, in most cases, for the 0.70. If that is why the model is in your stack, the successor is measurably worse at the property you bought it for, and the honest thing to do is measure whether 0.11 seconds of added delay costs you more than 15 points of task completion gains you. On a telephony deployment the answer is not obvious in either direction, and it depends on where in your call flow the agent is doing actual work.

The alias moved, and that is not a footnote

grok-voice-2.0 is deprecated as of the 14 October release, and grok-voice-latest now resolves to grok-voice-3.0. Anyone who deployed against the alias was moved onto the new model automatically, with no code change and nothing in their logs to indicate it.

Set that against the timeline and it becomes a pattern rather than an incident. On 5 August 2026 — about ten weeks before this release — grok-voice-latest was pointed at the model now called Think Fast 2.0, and developers who wanted to stay on the previous generation had to pin a versioned model ID before that date or absorb the change. That is two silent swaps in a single quarter, each one changing latency characteristics by more than a tenth of a second, with no version manifest and no deprecation window announced in advance.

For a chatbot that is fine. For a phone system with a regression suite written against a specific voice's timing, or an SLA that quotes an interaction latency, it means the alias is not the abstraction it appears to be. Pin the versioned ID in production, and treat grok-voice-latest as a staging convenience rather than a deployment target.

What did not change, and one thing that carries over

The upgrade is close to a drop-in, and it is worth being specific about why, because the alternative — GPT-Live-1, for instance — requires a WebRTC rewrite of the audio transport and a delegation pattern that has no equivalent here. Grok Voice 3.0 keeps the same realtime WebSocket endpoint, the same session.update parameters, the same event names, the same bearer-token handshake and the same short-lived ephemeral tokens for client-side connections. Voices are unchanged, and the custom-voice endpoint that trains a voice from a reference clip of up to 120 seconds produces ids that work across the realtime API and streaming text-to-speech exactly as before.

What carries over rather than improves is the transcription-side evidence. The previous generation was marketed heavily on word error rate — xAI reported evaluating thousands of short phrases across 24 languages and finding 1.5 to 2.0 times lower error than dedicated speech-to-text models including Deepgram Nova 3 and ElevenLabs Scribe v2, 1.4 times lower than its own predecessor, and roughly 10 times lower error in noisy conditions with background noise and telephony compression. Those are vendor figures and were never independently reproduced. Grok Voice 3.0's launch material does not restate them, which is neither a confirmation nor a retraction. If noisy-channel accuracy is load-bearing for your deployment, it is one of the few things you have to re-test yourself rather than read off a board.

Trying the new one without betting the phone line on it

The reason a version pin costs so little here is that nothing about Grok Voice 3.0 is hard to stand up side by side — and the reason a router is worth having in this stack is not the audio layer, which belongs to xAI alone, but everything the audio layer calls.

OrcaRouter does not serve the Grok Voice realtime endpoint and could not; a duplex audio WebSocket is a single-vendor surface. What we run is the text tier the voice model reaches for. A voice deployment is a voice model plus a classifier, a summariser, a retrieval step and a tool-calling backend, and that half is ordinary request-metered inference: 200-plus models behind one key, at provider list price with zero markup, so a vendor price cut anywhere in that layer lands on our side the same day rather than at renewal; automatic failover when a provider degrades, which for a live call is the difference between a retried lookup and a dropped caller; and a routing DSL that lets you compose several models into a single call and switch the mix on live traffic without touching the audio path.

That structure is what makes the pin cheap. Keep grok-voice-2.0 on production, point a second environment at grok-voice-3.0, and let the same backend models answer both — the comparison then measures the one variable you actually changed, and the code path underneath it is identical in both directions.

The verdict, in one line each

Upgrade if your agent does work rather than holds conversation. Fifteen points of τ-Voice and 59 points of arena preference at constant price is the clearest value-for-nothing trade in this category this year, and nothing in the integration has to move for it.

Pin and test first if your entire case for xAI's voice line was the 0.70-second start. That property is now 0.81, the model that beat it is still running, and it is deprecated rather than removed.

Ignore the naming. Grok Voice Think Fast 2.0, Grok Voice 2.0 and the API id grok-voice-2.0 all point at the same generation, and if you have been reading a leaderboard that lists only one of those names, you have been reading the right number under a label your own config file does not use.

Comparisons like this one are also where the failure modes show up: the documented AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive , where each entry carries its own sources, severity, confidence grade and disputed flag.