On October 14, 2026, xAI released Grok Voice 3.0, a speech-to-speech model for building voice agents, at $0.08 per minute — the same price as the Grok Voice 2.0 it replaces. If you run a small business, here’s the short answer: it’s a strong, fairly priced option for one narrow job — answering routine calls, booking appointments, and handling FAQs — but it is not a receptionist replacement, and the headline price understates what you’ll spend once telephony, tooling, and testing are in the picture. A typical 3-minute call costs about $0.24 in model time, and 500 such calls a month run roughly $120 for the model alone, plus phone and platform costs on top.
We have not tested the API ourselves. Every benchmark number below comes from xAI’s announcement or independent coverage of it, and vendor claims are labeled as such.
What Grok Voice 3.0 Actually Is
Grok Voice 3.0 is a speech-to-speech model. It listens to audio and answers in audio directly, in one model, instead of chaining three separate steps — speech-to-text, a text model, then text-to-speech. That shortcut matters because each handoff in the old chain added delay — and delay is what makes a robot phone call feel robotic.
You use it through xAI’s Voice Agent API. Your app opens a persistent WebSocket connection, streams the caller’s audio in, and gets spoken audio plus text back in real time. Inside that session, the agent can call tools: look up an appointment, check an order, search the web, or trigger a custom function you define. Per xAI’s docs, the API is compatible with OpenAI’s Realtime API, supports WebSocket, WebRTC, and SIP, offers 28 built-in voices plus custom voices cloned from a clip of up to 120 seconds, and covers 20 locales with automatic language detection.
One practical detail: agents pointed at the grok-voice-latest alias were switched to 3.0 automatically on launch day, with no code changes needed. That’s convenient, and it’s also a risk you’ll want to manage (more on that below). Grok Voice 2.0 is now deprecated. As of this writing, xAI has not published a shutdown date for 2.0 — but deprecation means no future improvements, so new builds should target grok-voice-3.0 directly.
The Benchmarks, Read Honestly
xAI’s headline claim is that Grok Voice 3.0 ranks first on Artificial Analysis’s Speech-to-Speech Index at 86.9, ahead of OpenAI’s GPT-Live-1 (Astra, Medium) at 83.2 and Google’s Gemini 3.8 Live (Extended Thinking, High) at 82.6. Grok Voice 2.0 scored 81.3. The index blends four components, so the composite hides as much as it reveals. Here’s the full table from xAI’s announcement:
| Benchmark | Grok Voice 3.0 | Grok Voice 2.0 | GPT-Live-1 (Astra, Medium) | Gemini 3.8 Live (Ext. Thinking, High) |
|---|---|---|---|---|
| AA Speech-to-Speech Index | 86.9 | 81.3 | 83.2 | 82.6 |
| Big Bench Audio (speech reasoning) | 98.5% | 97.2% | 90.1% | 97.7% |
| Speech Agent Arena (human preference, Elo) | 1070 | 1011 | 1048 | 990 |
| Task Success Rate | 94.8% | 94.6% | 87.4% | 89.1% |
| Full Duplex Bench (pauses, interruptions, turn-taking) | 97.0% | 95.1% | 94.9% | 91.9% |
Two things stand out. First, the Full Duplex score of 97% is the most meaningful number for a business phone line. It measures whether the model handles the messy mechanics of real conversation — pauses, interruptions, both people starting at once — without talking over your customer. Grok’s lead there (97.0% vs. 94.9% and 91.9%) is the gap most likely to show up in how natural your callers think the agent sounds.
Second, the human-preference lead is real but modest. In xAI’s published four-model comparison, Grok Voice 3.0 leads the Speech Agent Arena at 1070 Elo, 22 points ahead of GPT-Live-1. That’s a win, not a blowout, and preference is the most subjective measure in the set. Treat it as “callers in blind tests liked it slightly more,” not as proof your customers will love it.
The One Benchmark Where Grok Loses
Now the number xAI’s marketing leans on least. On τ-Voice (tau-Voice), a benchmark built from simulated customer-support calls with diverse accents, background noise, and interruptions, Grok Voice 3.0 scored 71.6%. GPT-Live-1 scored 74.5%. Gemini 3.8 Live scored 68.6%. xAI calls its result “comparable” to GPT-Live-1 — generous phrasing for finishing second.
This is the benchmark closest to what you’d actually deploy, and it’s the one where Grok doesn’t lead. The generational jump is still real: Grok Voice 2.0 managed only 56.5% on the same test. But read it plainly: on roughly 3 of 10 simulated support calls, the agent failed to complete the task. xAI’s counterweight is speed — 40% lower latency than GPT-Live-1, with responses in under one second. Artificial Analysis measured Grok’s time to first audio at 0.81 seconds versus 1.34 seconds for GPT-Live-1, which supports the claim.
xAI also reports 74% accuracy on EVA-Bench, ServiceNow’s benchmark for enterprise voice agents, calling it the highest reported result. No competitor scores were published alongside it, so treat that as a claim, not a ranking. Artificial Analysis also measured Grok’s cost per task on its Big Bench Audio run at $0.025, versus $0.030 for GPT-Live-1 and $0.055 for GPT-Realtime-2.1 (High) — an efficiency signal, not a bill you can budget from.
What a Voice Agent Actually Costs Per Call
The headline rate is $0.08 per minute, or $4.80 per hour. Here’s what that means in practice.
Cost Per Call
| Call length | Model cost at $0.08/min |
|---|---|
| 2 minutes (hours, directions, quick FAQ) | $0.16 |
| 3 minutes (typical booking call) | $0.24 |
| 5 minutes (booking + questions + reschedule) | $0.40 |
| 8 minutes (troubleshooting with tool lookups) | $0.64 |
A 3-minute call at $0.24 sounds almost free, and compared with a human’s time, it is. But model time is only one line on the bill.
Monthly Math at Realistic Volumes
Assume an average call of 3.5 minutes — bookings run longer than FAQ calls, and a few troubleshooting calls will drag your average up.
| Calls per month | Total minutes | Model cost |
|---|---|---|
| 200 (quiet shop, after-hours only) | 700 | $56 |
| 500 (busy salon, clinic, or trades business) | 1,750 | $140 |
| 1,000 (multi-location or heavy booking volume) | 3,500 | $280 |
| 2,000 (high-volume support line) | 7,000 | $560 |
Then add everything the headline rate doesn’t cover:
- Phone numbers and telephony. The Voice Agent API is the brain, not the phone line. You still need a phone number, SIP trunking or a telephony platform, and per-minute carrier charges. Budget for this separately; it can rival the model cost at low volumes.
- Text input charges. xAI’s pricing page also lists a separate $0.004 per text input charge, per launch coverage. Small, but it’s a second meter running.
- Your platform or developer. Unless you use a no-code voice-agent builder, someone has to wire the API to your booking system, calendar, or CRM. That integration — not the model — is usually the biggest first-year cost.
- Failed and abandoned calls. You pay for minutes whether or not the call ends in a booking. A confused caller who stays on the line for six minutes costs $0.48 and gives you nothing.
A realistic all-in budget for a small business running 500 calls a month is roughly $200–$400 per month once telephony and platform fees are included — not $140. Get a quote for the full stack before you compare vendors on headline rates.
Against a Human Receptionist
The U.S. Bureau of Labor Statistics puts the median receptionist wage at $18.27 per hour in May 2025, or about $38,010 per year full-time. Add payroll taxes, benefits, and overhead, and one full-time receptionist realistically costs roughly $4,000–$4,500 per month — and that’s one person covering one shift, who also greets walk-ins, handles mail, and does a dozen jobs a voice agent can’t.
That comparison cuts both ways. A voice agent at $200–$400 a month is dramatically cheaper if all you need is call handling. But a receptionist’s cost is fixed whether the phone rings or not, while the agent’s cost scales with minutes — and the agent can’t do the other 60% of the front-desk job. The honest framing isn’t “AI replaces your receptionist.” It’s “AI answers the calls your receptionist misses — after hours, on lunch breaks, and during rush periods — for a fraction of a second hire.”
For context, our breakdown of Nano Banana 2.1’s 50% price cut for creators walks through the same per-unit budget math for image generation: the per-unit rate is only the start of the bill.
Grok Voice 3.0 vs. OpenAI vs. Google
| Grok Voice 3.0 (xAI) | GPT-Live-1 (OpenAI) | Gemini 3.8 Live (Google) | |
|---|---|---|---|
| Headline voice price | $0.08/min | $0.05/min voice session | ~$0.005/min audio input + ~$0.018/min audio output |
| What’s included | Single speech-to-speech model; tools run inside the session | Voice layer only; backend reasoning model billed separately | Audio rates + thinking tokens for Extended Thinking |
| AA Speech-to-Speech Index | 86.9 | 83.2 | 82.6 |
| τ-Voice (support calls) | 71.6% | 74.5% | 68.6% |
| Full Duplex Bench | 97.0% | 94.9% | 91.9% |
| Arena Elo (human preference) | 1070 | 1048 | 990 |
| Time to first audio (Artificial Analysis) | 0.81s | 1.34s | Not published in the coverage checked |
| Architecture | One model does listening, reasoning, speaking | Voice frontend + swappable backend model | Live API with async tools, optional Extended Thinking |
| Telephony | SIP supported | Telephony supported | Via partner platforms (LiveKit, Agora, Vercel, others) |
The headline rates are not directly comparable — the most common mistake in voice-agent shopping. OpenAI’s $0.05 per minute looks cheaper than Grok’s $0.08 — until you add the backend model that does the actual reasoning, which OpenAI bills separately. Depending on the backend you pair with GPT-Live-1, your true per-minute cost can land above Grok’s flat rate. Google’s token-based pricing looks cheapest at roughly $0.023 per minute of audio alone, but Extended Thinking adds reasoning-token charges, and you pay for input and output audio separately.
Don’t pick on benchmarks alone. Prompt design, tool wiring, and handoff quality will move your results more than the model choice will.
Where a Voice Agent Beats a Human Receptionist
Voice agents win in specific situations. If your business matches these, a Grok-based agent is worth a pilot:
- After-hours calls. A plumber, HVAC company, or clinic that misses evening calls is losing bookings to whoever answers. An agent that books the 8 p.m. caller beats a voicemail every time.
- Peak-hour overflow. When three lines ring at once, the agent takes the overflow so nobody gets a busy signal or a 10-minute hold.
- Repetitive bookings and FAQs. “What are your hours?” “Can I move Thursday to Friday?” These calls are high-volume, low-judgment, and exactly what a voice agent handles well.
- Multilingual callers. Grok’s 20 locales with automatic language detection cover the common cases for most U.S. businesses. Google goes further at 97 languages, so if your customer base is heavily multilingual, weigh that.
Where It Fails, and When NOT to Use It
This is the section most launch coverage skips. A voice agent is a tool with a failure rate, and you should plan around it.
It will fail roughly 3 in 10 hard support calls. That’s what the τ-Voice score of 71.6% means in plain terms. Under unfamiliar accents, background noise, and mid-sentence interruptions, the agent gets things wrong or stalls. Your system needs a clean escape hatch: “Let me get a person for you” should be one sentence away, not buried in a menu.
Don’t let it move money unsupervised. xAI’s Starlink example has a single agent verifying customers, troubleshooting hardware, issuing credits, and shipping replacement equipment. That’s a showcase deployment with an enterprise team behind it. For your business, start with a hard rule: the agent can propose a refund, credit, or discount, but a human approves it. One wrongly issued $200 credit wipes out months of model savings.
Don’t use it for medical, legal, or safety advice. A clinic agent can book appointments and confirm prep instructions from your approved script. It should not interpret symptoms or triage emergencies. Same logic for legal intake: collect facts, don’t give advice.
Don’t use it as your only line for angry callers. People who are already upset want a human, and an AI that sounds cheerful while they’re furious makes it worse. Route repeat callers, refund requests, and anyone who asks for a manager to a person immediately.
Don’t let it take full card numbers. If a caller needs to pay, hand them to a secure payment link or a human. Reading card details to an AI agent creates a compliance problem you don’t need. If your business handles health information, ask xAI about its HIPAA eligibility and business associate agreement before you deploy — xAI lists SOC 2 Type II compliance and HIPAA eligibility with a BAA, per its documentation, but “eligible” means you still have to set it up correctly.
Don’t float on the grok-voice-latest alias in production. Auto-migration is convenient until a vendor ships a model that changes your agent’s behavior overnight — which is exactly what happened to every grok-voice-latest user on October 14. Pin grok-voice-3.0 explicitly, test new versions in staging, and promote them yourself.
Don’t skip disclosure. Tell callers up front they’re speaking with an AI assistant. It’s the honest move, it sets expectations (callers are more forgiving of a robot they know is a robot), and it keeps you aligned with disclosure rules that are tightening in several states. If you record calls, follow your state’s consent law for recording.
Don’t trust AI accuracy claims blindly. Our reporting on AI detectors wrongly flagging human-made music is a useful reminder: lab accuracy is not real-world accuracy. Benchmarks are a starting point. Your own test calls are the real evaluation.
The Starlink Claim: Read It Like a Buyer
xAI says Starlink call centers use Grok Voice to answer 20,000 customer calls every day across multiple languages, with roughly 70% resolved with no human in the loop, and one in four sales calls ending in a purchase before the caller hangs up. It also says Grok powers millions of conversations a day in Tesla vehicles, for hands-free tasks like checking a calendar or ordering a coffee.
Those are vendor-reported figures. No independent audit backs the Starlink numbers, and independent coverage of the launch has flagged exactly that. That doesn’t make the numbers false — Starlink is a real, high-volume deployment, and 70% containment is plausible for a support operation whose calls skew toward a few repeatable issues (outages, billing, hardware setup). But “resolved” is doing a lot of work in that sentence. A call can end without a human and still leave the customer unsatisfied, calling back the next day.
The useful takeaway isn’t the 70%. It’s that xAI is dogfooding its own model at scale before selling it to you. That’s a genuine positive signal — but it’s not a substitute for your own pilot data. Run 30 days, then measure your own containment, escalation, and callback rates. Those three numbers are your business case.
Setup Checklist: From Zero to a Live Voice Agent
If you decide to pilot Grok Voice 3.0, work through this list in order. Skip a step and you’ll pay for it in failed calls.
- Pick one call type. Booking, hours and FAQs, or order status — not all three. A narrow agent is a good agent. You can expand later.
- Map the happy path and five failure paths. Write out the ideal conversation, then the five ways it breaks: caller interrupts, wrong date, off-script question, angry caller, asks for a human. Script every one.
- Define what the agent may and may not do. Book, reschedule, cancel with confirmation? Yes. Issue refunds, change prices, access accounts without verification? No. Write the permissions down and enforce them in tool configuration, not just the prompt.
- Wire one tool at a time. Connect the calendar or booking system first. Test it with 20 fake calls before you add order lookup or payments.
- Pin the model version. Use
grok-voice-3.0, notgrok-voice-latest. Set a calendar reminder to review new versions quarterly. - Set up a staging number. Friends, family, and staff call it first. Then route 10% of real traffic. Then 50%. Then all of it. Never go from zero to 100% on launch day.
- Test the hard cases. Accents, background noise, a caller talking over the agent, a caller on a bad cell connection. If you only test in a quiet office, your first real-world week will humble you.
- Write the disclosure and handoff lines. “Hi, you’re speaking with [Business Name]’s AI assistant. I can help with bookings and questions — or I can get a person anytime, just say so.” Keep the handoff one sentence away, always.
- Set a budget cap and alerts. Cap monthly minutes in your platform and set an alert at 70% of budget. A misconfigured agent that stays on calls forever is a real way to burn money.
- Review calls weekly. Listen to 10 random calls and every escalated call. Fix the top failure, redeploy, repeat. The businesses that win with voice agents treat week one as the start of tuning, not the finish line.
The Bottom Line
Grok Voice 3.0 is the best voice model xAI has shipped, at the same $0.08-per-minute price as its predecessor, with the fastest first response in its class and the strongest full-duplex handling in the comparison set. For a small business, that adds up to a genuinely useful tool for after-hours coverage, overflow, and routine bookings — roughly $200–$400 a month all-in for 500 calls.
What it isn’t: a receptionist replacement, a money-moving autopilot, or a set-and-forget system. Its 71.6% score on simulated support calls is a reminder that hard calls still need humans, and Starlink’s 70% containment rate is a vendor-reported enterprise result, not a promise about your business.
Frequently Asked Questions
How much does Grok Voice 3.0 cost per call?
The model costs $0.08 per minute, so a 3-minute call costs $0.24 in model time. At an average of 3.5 minutes per call, 500 calls a month runs about $140 for the model alone; with telephony and platform fees, budget roughly $200–$400 per month. xAI’s pricing page also lists a separate $0.004 per text input charge.
Is Grok Voice 3.0 better than OpenAI’s GPT-Live-1?
On Artificial Analysis’s composite Speech-to-Speech Index, yes: Grok scored 86.9 versus 83.2 for GPT-Live-1. Grok also leads on Full Duplex Bench (97.0% vs. 94.9%) and human-preference Elo (1070 vs. 1048), and answers faster (0.81s vs. 1.34s to first audio). But GPT-Live-1 wins on τ-Voice simulated support calls (74.5% vs. 71.6%), and its $0.05-per-minute voice rate looks cheaper until you add the separately billed backend reasoning model. For most small businesses, deployment quality matters more than the benchmark gap.
Does Grok Voice 3.0 replace a human receptionist?
No — it covers part of the job. It answers calls after hours, handles overflow, and takes routine bookings and FAQs. It can’t greet walk-ins, handle mail, or exercise judgment on angry callers, refunds, or emergencies. Think of it as coverage for the calls your receptionist misses, not a replacement for the person.
Do callers know they’re talking to an AI?
Only if you tell them, so tell them. Open with a short disclosure — “You’re speaking with [Business Name]’s AI assistant” — and keep a human handoff one sentence away. Disclosure sets expectations, builds trust, and keeps you aligned with disclosure rules that are tightening in several states.
What happens to my existing Grok Voice 2.0 agent?
If your agent pointed at the grok-voice-latest alias, it was switched to 3.0 automatically on October 14, 2026, with no code changes. If you pinned grok-voice-2.0, it still runs, but 2.0 is deprecated with no published shutdown date. New builds should target grok-voice-3.0 explicitly, and you should re-test your prompts: the jump from 56.5% to 71.6% on τ-Voice means the new model behaves differently enough that prompts tuned for 2.0 may need adjusting.
Can Grok Voice 3.0 handle payments or refunds?
Technically it can call tools that do those things, but you shouldn’t let it act unsupervised. Start with a rule that the agent proposes refunds, credits, or discounts and a human approves them. Never have an agent read or repeat full card numbers — hand the caller to a secure payment link or a person instead.