Grok Voice 3.0 vs Gemini 3.8 Live: Google Sells Two Voice Models and Only One of Them Is in This Fight

G​oogle's answer to Grok Voice 3.0 costs $0.84 an hour. That is the Artificial Analysis cost-per-hour-of-input-audio figure for Gemini 3.8 Live in its standard configuration, and it is 5.7 times cheaper than the $4.80 an hour Grok Voice 3.0 bills at. It is also the wrong model to compare. The Gemini 3.8 Live that actually trades blows with SpaceXAI's new flagship — the one holding 82.6 on the Speech to Speech Index against Grok's 86.9 — is the Extended Thinking variant, and it costs $3.50 an hour. The real gap between these two models is 1.37×, not 5.7×, and picking the wrong row off the pricing table is the most expensive mistake available in this matchup.

This is the first time the category has had a genuine price-quality frontier rather than a leader and a cheaper alternative. Grok Voice 3.0 arrived on 14 October 2026 at $0.08 per minute, the same rate as Grok Voice 2.0, with 86.9 on the index and sub-second time to first audio. Gemini 3.8 Live landed on 15 September 2026 in two configurations that behave like two different products, one of which is a cost leader and one of which is a reasoning system that happens to talk.

The two Google models are not tiers of one another

Google's own framing separates them by job rather than by size. The standard model is described as built for scale and cost efficiency with fluid dialogue and visual grounding. The Extended Thinking variant is described as built for high-complexity tasks with multi-step reasoning that narrates its progress out loud while working.

The measured gap between them is not a quality gradient. It is a category break:

• Agentic task completion, τ-Voice — Gemini 3.8 Live Extended Thinking 68.6% vs Gemini 3.8 Live 30.1%. Same vendor, same release, same underlying family, 38.5 points apart.

• Speech to Speech Index — 82.6 for Extended Thinking at High reasoning effort vs 76.0 for the standard model.

• Cost per hour of input audio — $3.50 vs $0.84.

• Time to first audio — 1.35 seconds for Extended Thinking vs 1.18 seconds for the standard model, both measured by Artificial Analysis.

So when a procurement document asks whether Gemini 3.8 Live is cheaper than Grok Voice 3.0, there is no single answer. Against the standard model, xAI is nearly six times the price and 10.9 index points ahead. Against Extended Thinking, xAI is 37% more expensive on audio and 4.3 points ahead. Those are different decisions wearing the same model name.

Head to head, with the right Gemini

Set the standard variant aside for a moment and compare the two models that are actually competing for the top of the board. Every figure below is Artificial Analysis's own testing from 13 October 2026 for Grok Voice 3.0 and from the September board for Gemini 3.8 Live Extended Thinking at High effort.

• Speech to Speech Index — Grok Voice 3.0 86.9 vs Gemini 3.8 Live Extended Thinking 82.6. A 4.3-point gap that is real but not decisive.

• Agentic performance, τ-Voice — Grok Voice 3.0 71.6% vs Gemini 3.8 Live Extended Thinking 68.6%. Three points, and the closest thing on the board to a tie.

• Big Bench Audio speech reasoning — Grok Voice 3.0 98.5% vs Google's reported 97.7%. Both effectively at ceiling; this dimension has stopped discriminating.

• Speech Agent Arena preference — Grok Voice 3.0 at Elo 1070, the highest on the board. Google's model does not hold the arena.

• Price per minute of audio — Grok Voice 3.0 $0.08 in and out; Gemini 3.8 Live $0.005 per minute of audio input and $0.018 per minute of audio output.

• Time to first audio — Grok Voice 3.0 0.81 seconds; Gemini 3.8 Live Extended Thinking 1.35 seconds and the standard model 1.18 seconds on Artificial Analysis's measurement. Google's own materials publish no latency figure.

• Cost per completed task, Big Bench Audio — Grok Voice 3.0 $0.025 per task. Neither Google variant has a published figure.

• Availability — Grok Voice 3.0 is priced and live with a deprecation notice on its predecessor; Gemini 3.8 Live is a preview with no general-availability date, no published uptime figure and no rate Google has promised to hold.

• Language coverage — Gemini 3.8 Live switches across 97 supported languages mid-conversation; xAI's speech-to-text tier documents 38-plus languages.

• Visual input — Gemini 3.8 Live processes near real-time visual input; Grok Voice 3.0 has no comparable visual grounding.

The one dimension where Google's silence is the story

Notice the asymmetry in that list. Grok Voice 3.0 starts speaking in 0.81 seconds. Gemini 3.8 Live Extended Thinking starts in 1.35 and the standard model in 1.18, both on Artificial Analysis's measurement — which makes the non-reasoning Gemini slower to begin a turn than the reasoning one, an oddity that is almost certainly the effect of the reasoning variant narrating its progress out loud while it works rather than of a genuinely faster path.

Google's own launch material publishes no latency figure for either model, only partner language about impressive latency. That matters because time to first audio is the dimension where voice products succeed or fail in the eyes of the people using them, and it is the dimension Google chose not to put a number on. If your product is telephony, where 0.81 seconds is the difference between a conversation and a delay, a competitor's measurement of your vendor is not the same as your vendor's commitment. Test before you commit, and watch the general-availability announcement to see whether a figure appears alongside it.

The visual grounding cuts the other way and is unambiguously Google's. A model that can process near real-time video can resolve the things audio cannot: what "this" refers to, whether the person asking is looking at the thing they are asking about, whether the noise on the line is a machine that can be seen. Grok Voice 3.0 is an audio model with tools. That is a genuine capability boundary, not a configuration difference, and no amount of index score closes it.

What the price difference actually buys

Run the arithmetic on a real workload. Ten thousand minutes of voice a day, thirty days a month, three hundred thousand minutes.

• Grok Voice 3.0 — $0.08 per minute, so $24,000 a month.

• Gemini 3.8 Live Extended Thinking — priced from $3.50 per hour of input audio, roughly $17,500 a month for the same volume, with a public rate card that charges input and output separately.

• Gemini 3.8 Live standard — roughly $4,200 a month, on a model that completes 30.1% of τ-Voice scenarios.

That middle row is the whole decision. For roughly $6,500 a month less than xAI's flagship, Google sells a model 4.3 index points behind it on a preview status — close enough that any team whose agent is doing genuine multi-step work should be running both against their own recordings rather than reading a leaderboard. The bottom row is a trap unless your voice interface is conversational and your task completion happens elsewhere in the stack, which for a large share of deployed voice products it is.

There is one more measure worth carrying into that spreadsheet, because it partly reverses the picture. Artificial Analysis's per-task measurement on Big Bench Audio puts Grok Voice 3.0 at $0.025 to complete a task against $0.030 for G​PT Live 1 in its Astra medium configuration. Cost per completed task, not cost per hour, is the number that survives contact with a budget meeting — a cheaper model that needs a second attempt is not cheaper. Neither vendor publishes that figure for the Gemini pair, which is another blank where the cheaper option has not been measured on the axis that matters.

Where a router helps and where it does not

Be precise about this, because the reflex answer is wrong. OrcaRouter does not serve either realtime endpoint. Grok Voice 3.0 comes from xAI's own API and Gemini 3.8 Live from Google's, and no router can interpose itself on a duplex audio WebSocket belonging to one vendor. If someone tells you their platform lets you fail over between these two mid-call, they are describing something that does not exist.

What is routable is everything around them. A voice deployment is a voice model plus a text stack — the classifier that assigns a disposition code after the call, the summariser that writes the CRM note, the retrieval that answers the question the voice model had to hand off, the model reading the transcript for compliance. That layer is ordinary request-metered inference, it is where a meaningful share of the per-call cost accumulates, and it is the layer we run: 200-plus language models behind one key at provider list price with zero markup, automatic failover when a provider degrades, and a routing DSL that lets you compose several models into one call rather than picking one in advance. Because list price passes through unchanged, a vendor price cut on any of those models is live on our side the same day. The audio minute you pay Google or xAI for is the same minute either way; the reasoning behind it is the part you get to choose.

The decision rule

Three answers, and they follow from the components rather than the index.

Pick Grok Voice 3.0 if your workload is agentic, latency-sensitive and going into production this quarter. It holds the composite, it holds the arena, it holds the fastest published start-to-speak time among the frontier models, and it is priced and supported with a deprecation path rather than a preview notice.

Pick Gemini 3.8 Live Extended Thinking if you need visual grounding, if you need 97 languages rather than 38, or if you are running enough voice minutes that $1.30 an hour of difference matters more than 4.3 index points — and accept that you are building on a preview with no uptime commitment.

Pick the standard Gemini 3.8 Live only if you have measured your own traffic and confirmed that the multi-step part of your calls is already handled outside the voice model. On paper it is the cheapest option in the category by a wide margin. On the τ-Voice axis it fails roughly seven of every ten scenarios, and nothing in the index tells you that until you look.

What would change this page: a general-availability date and a latency figure from Google, which together would make the cheaper option the safer one; or a price move from xAI, which is the one lever its competitor cannot match, given that the flagship rate has now held constant across a full model generation.

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.