Grok Voice 3.0 vs GPT-Live-1 vs Gemini 3.8 Live: benchmarks, real per-minute costs, and what I'd put on a phone line

On 14 October xAI released Grok Voice 3.0, and on Artificial Analysis' independent speech-to-speech index it is now the top model, ahead of OpenAI's GPT-Live-1 and Google's Gemini 3.8 Live. It costs $0.08 a minute, the same as the version it replaces.

I build AI voice agents that answer and place real phone calls, so a new leader in speech-to-speech is directly relevant to what I ship. Benchmarks and price lists only tell you so much, though. Here is what each of the three actually offers, what it costs per minute and per task, and which one I'd put on a phone line.

The benchmarks

These are Artificial Analysis results, as published in xAI's announcement — xAI chose which competitors to show, so read it as the strongest version of Grok's case:

Grok leads on five of six. The one it doesn't lead is the one I care about most: τ-Voice, which measures whether a voice model can complete a multi-step task using tools — look up the order, check the policy, make the change. GPT-Live-1 is ahead there. xAI's write-up calls the two "comparable", and three points is close, but it is the benchmark closest to what a customer-service agent does all day.

The other number worth noticing is Full Duplex Bench: 97% on handling interruptions, back-channels and turn-taking. On a phone line that matters more than raw reasoning — callers talk over the agent, say "mm-hm", and change their minds mid-sentence. A model that handles that badly feels broken no matter how smart it is.

Latency: xAI says time to first audio is under one second and 40% lower than GPT-Live-1 at similar τ-Voice accuracy. That's a vendor claim rather than an independent measurement, so treat it as a hypothesis to test on your own network — for phone calls, the carrier leg and your own server location add more than the gap between these models. My latency budget calculator shows where the rest of the time goes.

The pricing models are completely different

This is the part comparison tables flatten, and it's the part that decides your bill.

Three things follow:

Grok is the most predictable. One number per minute, model included. For a business that budgets per call, that is genuinely valuable — you can price a service before you've run a single call. The flip side: you pay the same for an easy minute as a hard one, and you should check how connected-but-silent time (hold music, a caller looking for an order number) is counted on your traffic.

GPT-Live-1's $0.05 is a floor, not a price. OpenAI's pricing page is explicit that backend model and tool usage is charged separately. The useful comparison isn't $0.05 against $0.08 — it's the all-in cost per call, which you only get by measuring.

Gemini 3.8 Live is the cheapest on paper — a few cents a minute of audio — but per-token billing with thinking charged at the output rate makes it the hardest to forecast. Longer reasoning means a bigger bill, and you won't know your number until you've run real calls.

The independent number that cuts across all three is cost per task. On Artificial Analysis' Big Bench Audio run, quoted by xAI: Grok Voice 3.0 $0.025, GPT-Live-1 $0.030, GPT-Realtime-2.1 $0.055. Grok's list price per minute is higher than GPT-Live-1's, yet it came out cheaper per answered question — the per-minute rate isn't the whole story. That's the right way to compare voice models — cost per completed conversation, not cost per minute — and it's also why your own measurement beats any table, including this one.

For scale: 10,000 four-minute calls a month is 40,000 minutes. Grok Voice 3.0 is $3,200 of AI before telephony; GPT-Live-1 starts at $2,000 before its backend model. My voice-agent cost calculator now includes both, alongside a cascaded pipeline and the carrier costs that usually dominate in South Asia, Africa and Latin America.

Building with it

The engineering details matter more than the benchmark for anyone actually shipping:

• It speaks OpenAI's Realtime protocol. xAI's docs say the API is compatible with the OpenAI Realtime API, and switching usually means changing the base URL and key (with some differences in events). That makes Grok cheap to try if you already have a Realtime integration, and keeps you portable if you don't.
• Telephony is first-class. SIP, WebSocket and LiveKit connections, plus native G.711 μ-law and A-law at 8 kHz — the codecs phone networks actually use — so no transcoding step adds latency or artifacts.
• Tools run server-side. Web search, X search, file search over your collections and remote MCP servers run automatically; only your own custom functions come back to your code. Less round-tripping, fewer moving parts.
• Reasoning is adjustable with a reasoning.effort setting, so you can trade depth for speed per agent.
• Voices and cloning: built-in voices, plus custom voices cloned from a short reference clip.

Where it falls short for my market

xAI lists 20+ languages with automatic detection, including Hindi and several Arabic variants. Urdu isn't on the list. The docs say other languages may work "with varying accuracy", and spoken Hindi and Urdu overlap heavily — but "overlaps heavily" isn't good enough for a bank, a clinic or a government helpline in Pakistan, where vocabulary, politeness forms and the way numbers and amounts are read out all differ.

I've spent a lot of time on Urdu voice systems — getting a model to answer in clean, natural Urdu rather than a Hindi–English mix, and reading amounts the way people actually say them. For Urdu-first phone lines I'd still test any speech-to-speech model hard against a cascaded pipeline, where you choose the speech recognition, the language model and the voice separately and can fix each one. That trade-off — integrated speed versus per-stage control — is the real decision, and it's language-specific.

What I'd pick

English (and the listed languages), customer-facing calls, predictable budget: Grok Voice 3.0. Best measured conversational behaviour, a flat rate you can price against, and the scale evidence xAI cites — Starlink's call centres answering about 20,000 calls a day, roughly 70% without a human (xAI's figures). The Realtime compatibility makes it a low-risk trial.

Agents that do multi-step work with tools — booking changes, account updates, troubleshooting flows: test GPT-Live-1 alongside it. It leads on τ-Voice, and that's the benchmark that maps to those calls. Budget on measured all-in cost, not the $0.05.

Lowest cost at high volume, and you can tolerate forecasting work: Gemini 3.8 Live, with a cap on reasoning effort and a month of measured calls before you commit to a price.

Urdu, Pashto or other languages off the list: a cascaded pipeline, or a careful bake-off with real callers before any speech-to-speech model goes live.

Whichever you pick, keep the model behind your own interface. The top of these rankings has moved every few months this year, and the useful position is being able to switch providers in an afternoon, with your own test calls to tell you when it's worth doing.

References: xAI: Introducing Grok Voice 3.0 (14 October 2026; benchmark table from Artificial Analysis, cost-per-task figures, Starlink and Tesla deployment numbers) · xAI speech-to-speech docs (Realtime compatibility, languages, codecs, tools) · xAI models and pricing · OpenAI API pricing (GPT-Live-1 and GPT-Realtime-2.1) · Gemini API pricing (Gemini 3.8 Live). Prices checked against each provider's own pages on 19 October 2026.