Grok Voice 3.0 Tops the Speech to Speech Index at 86.9 — and Gives Back the Latency Record

The number everyone will quote is 86.9. That is where Grok Voice 3.0 landed on the Artificial Analysis Speech to Speech Index when the board was re-run on 13 October 2026, five and a half points clear of the model it replaces, Grok Voice 2.0, which sat at 81.3. It is also clear of GPT-Live-1 in its Astra medium configuration at 83.2 and Gemini 3.8 Live Extended Thinking at 82.6, which makes it the first xAI voice model to hold the top row outright rather than share a claim to it. SpaceXAI announced it on 14 October, the same day the API release notes changed, and at the same price per minute as the model it retires.

The number nobody will quote, because xAI did not put it in the headline, is 0.70. That is Grok Voice 2.0's time to first audio in seconds, and it is faster than Grok Voice 3.0's 0.81. The company that has spent a year selling sub-second turn-taking now fields a flagship that starts speaking 0.11 seconds later than its own outgoing model. It is still roughly 40% quicker off the mark than the GPT-Live-1 configuration Artificial Analysis measured at 1.34 seconds, so the marketing claim survives a straight reading. It just does not survive a comparison to the previous version.

What actually shipped, and what it costs

Grok Voice 3.0 is a speech-to-speech model served over the same realtime WebSocket endpoint as its predecessor, with the same session parameters and the same event names. There is no new SDK and no migration document, because there is nothing to migrate. The model string changed and the behaviour behind it changed.

• Price — $0.08 per minute of audio, or $4.80 per hour, plus $0.004 per text input. Identical to Grok Voice 2.0. This is a capability update at constant price, not a repricing.

• Alias behaviour — grok-voice-latest now resolves to grok-voice-3.0. Anyone who deployed against the alias was moved onto the new model automatically, without a code change and without a version pin.

• Deprecation — grok-voice-2.0 is flagged deprecated. The upgrade path xAI documents is to set the model field to grok-voice-3.0 or leave the alias in place.

• Endpoint and auth — unchanged. Same realtime URL, same bearer-token handshake, same short-lived ephemeral tokens for client-side connections.

• Voices — the built-in voice set and the custom-voice endpoint are unchanged; custom voices are trained from a reference clip of up to 120 seconds and the resulting voice_id works across the realtime API and streaming text-to-speech.

That last list is the reason the alias move matters more than it sounds. Grok Voice 2.0 was only made the alias target on 5 August 2026 — about ten weeks before this release. A team that built against grok-voice-latest in August has now been through two silent model swaps in a single quarter, which is a fine property for a chatbot experiment and a less fine one for a phone system whose regression suite was written against a specific voice's timing.

The index, component by component

Composite indices hide the argument, so it is worth pulling Grok Voice 3.0 apart. Every figure below comes from Artificial Analysis's own testing, dated 13 October 2026.

• Speech reasoning, Big Bench Audio — 98.5%, up from 97.2% for Grok Voice 2.0.

• Conversational dynamics, Full Duplex Bench — 97.0%, up from 95.1%.

• Speech Agent Arena preference — Elo 1070, up from 1011, and the highest on the board.

• Task success — 94.8%, up from 94.6%. This is the smallest movement of the five and worth noting as such: on raw task completion the two models are, within measurement, the same model.

• Agentic performance, τ-Voice — 71.6%, up from 56.5%. This is the largest single move in the release, a 15-point jump on the one component that measures whether an agent can actually complete a multi-turn job rather than merely sound competent doing it.

• Time to first audio — 0.81 seconds, against 0.70 for Grok Voice 2.0.

Reading those together tells a more useful story than the composite does. The 5.6-point index gain is not evenly distributed: nearly all of it is agentic performance and arena preference, with reasoning and duplex handling nudging upward and raw task success flat. Grok Voice 3.0 is a better agent, not a better conversationalist. If your voice product is a support line that books appointments and checks order status, that is precisely the delta you want. If it is a companion or a roleplay front-end, the index gain is largely beside the point, and you have paid a latency penalty for it.

The one test where the leader is not the leader

τ-Voice is the component xAI should be talking about, and the one where it still trails. At 71.6%, Grok Voice 3.0 is up enormously on its predecessor's 56.5% — and still behind GPT-Live-1 configured with Astra at medium reasoning, which Artificial Analysis puts at 74.5%. Gemini 3.8 Live sits at 68.6% on the same test.

So the shape of the category after this release is: xAI holds the composite and the arena; O​penAI holds the agentic task-completion test; Google holds neither and is cheaper than both. None of those three statements contradicts the others, and any article that reports only the first is reporting a ranking rather than a result.

The vendor claims that no board covers

The performance claims come with a second layer of numbers that Artificial Analysis does not measure and cannot, because they describe production traffic inside one company's own products. These are xAI's figures, reported by xAI, with no independent verification available and none likely:

• Tesla — Grok Voice is said to power millions of conversations per day in vehicles, handling calendar and email actions through connectors and food ordering through Grok Bot's voice mode.

• Starlink support — 20,000 customer calls per day across multiple languages, with roughly 70% resolved without a human and about one in four sales calls ending in a purchase. The same agent is described as handling identity verification, hardware troubleshooting, account credits and replacement shipments in one flow.

• ServiceNow EVA-Bench — a rounded 74% accuracy figure, reported as the highest on that benchmark and the best when accuracy and customer experience are scored jointly. EVA-Bench covers customer service at 15 tools and 50 scenarios, IT service management at 59 tools and 80 scenarios, and HR at 47 tools and 83 scenarios.

Treat the first two as a description of intent rather than as evidence. They are the strongest possible proof that a voice model works — a company routing its own revenue calls through it has skin in the game — and simultaneously the weakest possible evidence for a reader, because none of it is auditable from outside and none of it separates the model from the prompts, retrieval and escalation policy wrapped around it. The 70% containment figure in particular is a system-level number wearing a model-level costume.

What a price that did not move tells you

Holding the rate at $0.08 per minute is the most strategically interesting thing in the release. The obvious read is competitive pressure: Google has spent the year positioning Gemini Live as the cheap option in this category, and the prior generation of xAI's voice line had already absorbed a 60% price increase when the Think Fast line moved from $3.00 to $4.80 per hour. Raising the flagship rate again was not available.

The second-order effect is on anyone modelling cost. Artificial Analysis's per-task measurement puts Grok Voice 3.0 at $0.025 to complete a Big Bench Audio task, against $0.030 for GPT-Live-1 Astra medium and $0.055 for GPT-Realtime-2.1 High. Cheaper per task and more accurate per task is the combination that ends procurement debates, and it is the first time in this category that a single model has held both ends.

Where a routing layer enters this picture is narrower than it looks, and it is worth being exact. OrcaRouter does not serve the Grok Voice 3.0 realtime endpoint — that comes from xAI's own API, as does the Gemini Live and GPT-Live-1 equivalent. What we do run is the text side: 200-plus language models behind one key at provider list price with zero markup, which is where the transcription cleanup, the summarisation after the call, the classification of a transcript into a disposition code and the tool-calling logic that the voice model delegates to tend to live. Those are ordinary request-metered models, they are exactly the workload a router is for, and because list price passes through unchanged a vendor price cut on any of them is live on our side the same day rather than at your next renewal.

What is still unmeasured

Three things, and they are what should decide whether you move now or pin.

The first is long-session stability. Every figure above comes from short, scored interactions. Nothing published covers what a 20-minute call does to error rates, and the interesting failure mode for a duplex model — accumulating drift, or a turn-taking policy that degrades under barge-in — only shows up over minutes, not turns.

The second is the noisy-channel claim. xAI's predecessor generation was marketed on word error rate in telephony conditions, and that is the most falsifiable claim in the entire launch. Anyone with a call recording corpus and an afternoon can check it.

The third is whether the 0.81-second figure holds under load. Time to first audio is measured on a quiet system. Production is not a quiet system.

Until those land, the honest summary of Grok Voice 3.0 is this: it is the best voice model on the public board by a margin that is real but concentrated in one dimension, it costs the same as the model it replaces, it is slower to start speaking than that model, and it moved every existing grok-voice-latest deployment onto itself without asking. The first three are reasons to test it. The fourth is a reason to have a version pin in your config before you do.

A voice-model release is also the moment to check what has gone wrong before: the documented AI incidents and agent failures around 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.