Grok Voice 3.0 review: benchmarks, pricing, and where GPT-Live-1 still wins

What Grok Voice 3.0 actually is

SpaceXAI announced Grok Voice 3.0 on 14 October 2026 and calls it a speech-to-speech model: one model that hears a caller, reasons, calls tools and speaks back, instead of chaining speech-to-text into a text model and then back into text-to-speech. That single-pass design is the whole pitch — a chained pipeline loses tone at every handoff, which is why a caller can sound annoyed twice before the bot notices.

On the name, because you will see this vendor written three ways: it is xAI, the lab that shipped the earlier Grok models, and it now brands itself SpaceXAI. The site and docs carry SpaceXAI branding, the footer reads "SpaceXAI LLC", the documentation is titled "SpaceXSI Docs" and the social handle is @SpaceXSI — while the GitHub organisation is still xai-org and the API is still served from api.x.ai. The product pages on this site use SpaceXAI where we are quoting the vendor's own branding, and xAI where we mean the company.

You do not buy a box or install anything. You open a WebSocket to the Voice Agent API, send a session.update with the agent's instructions, voice and tools, then stream audio both ways until the call ends. SIP, WebSocket and WebRTC are all supported transports, and there are official tester apps for iOS, the browser and Twilio telephony in SpaceXAI's cookbook repository.

The endpoint, session parameters and event names did not change between 2.0 and 3.0. That is what makes the alias swap invisible — and also why an agent can change behaviour without anyone touching the code, which we come back to in what changed.

The benchmark table, and who actually produced it

Every score below comes from Artificial Analysis, dated 13 October 2026, and was published inside SpaceXAI's launch post. That matters twice over: it is a third-party measurement, and it is a third party's measurement re-published by the vendor. We are quoting the table, not re-running it.

Four of those rows measure four different things, and collapsing them into one "best model" number hides the only decision that matters. Speech reasoning is whether the model understands what was said. Arena Elo is which model people preferred talking to. Task success is whether the job actually got done in a live call. Full Duplex Bench is pauses, barge-in and not talking over the caller. tau-Voice is the one that adds tools and a policy document and only scores when the end state is exactly right — which is why it is the row a support team should read first.

SpaceXAI also reports 74% on EVA-Bench, ServiceNow's open benchmark of 200-plus enterprise voice scenarios, and quotes ServiceNow's product director describing a Pareto frontier on accuracy and customer experience. That is a vendor-reported figure from a named third-party benchmark; we have not seen an independent re-run.

What actually changed from Grok Voice 2.0

The headline is not the index. It is tau-Voice going from 56.5% to 71.6% — a 15-point jump in finishing tool-using support calls, after barely moving between the 1.0 and 2.0 generations. Everything else moved by a few points:

• Overall index: 81.3 to 86.9.
• tau-Voice: 56.5% to 71.6%. The real behavioural change, and the one most likely to disturb a prompt tuned against 2.0.
• Full duplex: 95.1% to 97.0%. Smoother pauses and interruptions.
• Speech reasoning: 97.2% to 98.5%. Close to saturated.
• Time to first audio: 0.70s to 0.81s. 2.0 was the faster model to start speaking; 3.0 gives about a tenth of a second back.
• Price and API: unchanged. Same $0.08 per minute, same endpoint, same session parameters.

Read together, 2.0 was the speed release and 3.0 is the brains release: it spends a fraction of a second more before it starts talking and buys a much better chance of finishing the task. For a support line that is usually the right trade, and 0.81 seconds is still well under a second.

The operational catch is the alias. If your agent sends model=grok-voice-latest, it moved to 3.0 on 14 October with no deploy on your side. A model that completes more tasks also gets there by different routes — when it yields the turn, when it escalates, when it calls a tool — so a prompt tuned against 2.0 should be re-tested even though nothing on your end changed.

Grok Voice 3.0 versus GPT-Live-1: the honest scoreboard

This is the comparison worth reading carefully, because both vendors are telling the truth about different rows.

SpaceXAI's own framing of the tau-Voice row is "comparable with GPT-Live-1 (Astra), but at 40% lower latency". Comparable is doing real work in that sentence: on the row where an agent has to actually resolve a support call, GPT-Live-1 is ahead, and Grok's case is that it gets there faster and cheaper. If your agent's job is answering questions and handing off, latency and cost probably win. If it has to close out the call itself, the tau-Voice gap is the number to watch.

One discrepancy worth flagging, because you will see both claims in the wild: SpaceXAI's own voice landing page has carried a "#1 Tau Voice Leaderboard" line, while the Artificial Analysis tau-Voice run places Grok Voice 3.0 second behind GPT-Live-1 (Astra). Independent coverage of the launch noticed the conflict and attributed it to different leaderboards. Treat vendor "#1" badges as needing the specific board named — including the ones on this site.

For the OpenAI side of this comparison, start from the GPT-Live overview, the GPT-Live API guide or the GPT-Live vs Realtime API breakdown. This page deliberately does not restate the OpenAI pricing, because a stale second-hand number is worse than a link to the page we keep current.

Pricing and how you get access

Grok Voice 3.0 is sold per minute through the Voice Agent API, and the price did not move with the model:

• $0.08 per minute — the same rate as Grok Voice 2.0, which SpaceXAI also states as $4.80 per hour.
• $0.004 per text input — a separate line item in SpaceXAI's voice API pricing.
• Access: an xAI/SpaceXAI API key from the console, or an ephemeral token for browser and mobile clients so the key never ships to the client.
• Starting points: the API playground and the cookbook tester apps — a Swift iOS sample, a browser WebSocket agent, a WebRTC agent and a Twilio telephony agent.

Because the API is unchanged, upgrading an existing 2.0 agent is a one-line change: set model to grok-voice-3.0. If you were already on the alias, it is a zero-line change.

Do not read "$0.08 per minute" as your total bill. It is the audio rate; text inputs are billed separately, and any tools your agent calls add their own cost. Budget from your own call durations rather than from the per-task averages above, which come from one specific benchmark run.

It is already running at scale

Most model launches are demos. This one shipped with production numbers attached, and they are the most useful part of the announcement because they are about what happens after the caller stops being polite:

• Tesla vehicles. SpaceXAI says Grok Voice handles millions of conversations a day in Teslas, with Connectors for simple requests and Grok Bot for longer tasks such as ordering coffee.
• Starlink support. 20,000 customer calls a day across multiple languages, roughly 70% resolved with no human in the loop, and about one in four sales calls ending in a purchase before the caller hangs up.
• What the agent does: verifies customers, troubleshoots hardware, issues credits and arranges replacement equipment — a tool-calling workload, not a chat demo.

Those figures are SpaceXAI's, and they are the reason to take the tau-Voice jump seriously even though the model scores second on it: several thousand live calls a day is a much harder test than any benchmark, and it is the vendor's own operational claim rather than a measurement we can reproduce.

What this page could not verify

• We have not run it. No calls, no latency measurements, no agent rebuilt on 3.0 for this page. Every score is reported by SpaceXAI or Artificial Analysis and quoted here as reported.
• "World's most intelligent voice model" is a vendor claim, not a finding. The defensible version is narrower: first on one composite index, second on tau-Voice.
• Benchmarks are not your workload. tau-Voice simulates support calls; Arena Elo is human preference after the fact. Neither promises your prompt, your tools and your caller demographics behave the same way.
• Time to first audio is not task latency. It measures when speech starts, not how long the caller waits for the job to finish.
• No retirement date for Grok Voice 2.0 has been published, so "deprecated" here means exactly that and no more.
• Free-trial terms were not checked. The launch page offers a free entry point; we have not verified what it currently includes.

If any of those gaps matter to your decision, the honest next step is a short bake-off: run your own last fifty calls against both models with the same prompt and compare completion, escalation and cost per resolved call. That is the only number that will match your business.

Sources

• SpaceXAI — Introducing Grok Voice 3.0: launch date, benchmark table, EVA-Bench figure, Starlink and Tesla production numbers, and the $0.08 per minute price.
• Artificial Analysis — Speech-to-Speech leaderboard: the independent index and arena results the table above quotes.
• SpaceXAI docs — Speech to Speech: model IDs, endpoint, session parameters, transports and tester apps.
• SpaceXAI docs — Release notes: the 14 October entries for the 3.0 launch and the Grok Voice 2.0 deprecation.
• SpaceXAI docs — Voice API pricing: per-minute rate and the separate per-text-input charge.
• ServiceNow — EVA-Bench: the open enterprise voice-agent benchmark behind the 74% figure.
• iClarified — Grok Voice 3.0 launches: independent coverage carrying the 0.81s versus 1.34s first-audio comparison.
• eesel.ai — Grok Voice 3.0: what it is, what changed: the independent write-up that flagged the tau-Voice leaderboard conflict and the generation-by-generation numbers.

Related on this site