Grok Voice 3.0 Costs $0.08 a Minute, but the Real Call Bill Is Bigger
xAI kept the price flat and raised task scores, but latency slipped. Here is what a real call costs once carrier, tools, and silence are counted.
Jump to 7 sections
xAI released Grok Voice 3.0 on October 14, 2026, keeping the price at $0.08 per minute. That rate covers speech-to-speech only. Phone numbers, tool calls, text inputs, and silence on the line are billed on top, so a four-minute support call costs closer to $0.35 than $0.32.
Voice agents are the AI product where the pricing page and the invoice disagree most. A per-minute rate sounds simple until you add the carrier, the lookups, and the dead air. Model Drop tracks launches and the tools built on them, and this brief covers what xAI shipped, what its numbers say, and what a realistic call costs.
It is written for developers and product teams evaluating a voice model for support, sales, or in-app assistants.
What xAI Shipped on October 14
According to AI Weekly's launch alert, xAI released Grok Voice 3.0 on October 14, 2026, at $0.08 per minute, unchanged from Grok Voice 2.0. The model is a speech-to-speech system, meaning it listens to audio and replies in audio without converting to text in between. It is aimed at high-volume conversational calls.
The launch post, as relayed by the same alert, claims a score of 86.9 on the Artificial Analysis Speech-to-Speech Index, described as the top score. It also cites 97% on Full Duplex Bench, 71.6% on tau-Voice, and 74% on EVA-Bench. ServiceNow's Nitin Aggarwal is quoted as saying the model sets a new Pareto frontier on EVA-Bench.
One change affects existing users directly. TechEnclave reports that customers using the latest-model alias were moved to Grok Voice 3.0 automatically. That changes live integrations without a code change on your side.
Model Drop has not tested Grok Voice 3.0 independently. Where this brief cites performance, it names the source.
Is Grok Voice 3.0 Actually Better Than Version 2.0?
Grok Voice 3.0 is better on task scores but slower to respond. Artificial Analysis scores it 86.9 on its speech-to-speech index, up from 81.3 for Grok Voice 2. Response latency, however, rose to 0.81 seconds from 0.70 seconds.
TechEnclave's summary puts it plainly: the tests show a mixed trade-off rather than an improvement on every measure. Developers should evaluate task scores, latency, and real usage charges separately.
Latency matters more in voice than in text. A pause longer than about a second makes a caller feel they are talking to a machine. A tenth of a second may not be noticeable, but it adds up across a call with many turns.
The comparison with OpenAI is also mixed. xAI says its response latency is sub-second and 40% lower than GPT-Live-1. A report relayed by AI Weekly says latency is 0.81 seconds, a regression from version 2.0, and that the model trails GPT-Live-1 on a customer-service benchmark. Both statements trace back to different measurements.
That is a reminder to read launch claims carefully. Our guide to reading a model launch announcement without getting sold lists the questions to ask, starting with who ran the test.
What Does a Real Grok Voice Call Cost?
A real call costs somewhat more than the per-minute rate suggests. The base price is $0.08 per minute with no seat fee, platform fee, or minimum. Text input into a voice session adds $0.004 each, and tools bill separately.
An eesel pricing breakdown works through a sample. The author, whose company sells a competing helpdesk product, assumes a four-minute call on a Twilio local number with three file searches and one text input. The pieces add up like this:
| Line item | Cost |
|---|---|
| Grok talk time, 4 minutes at $0.08 | $0.32 |
| Phone carrier | $0.0136 |
| Three file search calls | $0.0075 |
| One text input | $0.004 |
| Total | About $0.345 |
That works out to about $0.086 per talk minute, roughly 8% above the sticker rate. A toll-free number raises the total to about $0.384.
xAI does not sell phone numbers through its API, so you bring your own carrier. Speech-to-text and text-to-speech are separate APIs, billed at $0.10 to $0.20 per hour and $15 per million characters. Built-in search tools cost $2.50 to $5 per thousand calls, and your own function tools carry no xAI fee.
Scale changes the picture. At 500 four-minute calls a day for 30 days, the same breakdown estimates about $5,178 a month, with talk time making up around 93% of it. For a broader view of how voice products price their time, our comparison of text-to-speech models by latency, naturalness, and cost is a useful companion.
Why Does Dead Air Matter So Much?
Dead air matters because the clock runs whether anyone is talking or not. The eesel breakdown advises budgeting as though every open minute is billed, since the documentation does not exclude silence.
The arithmetic is simple. At 500 calls a day, 20 extra seconds of silence per call adds up to 300,000 seconds a month. That is 5,000 minutes, or about $400 at $0.08 per minute. The author notes that this exceeds the combined cost of carrier fees, tools, and text inputs in the same example.
The suggested fixes are practical. Set an idle timeout so the session ends after a pause, and use the hangup endpoint when the task is done. Send caller context as a single text input at the start instead of many small ones. Answer from your own uploaded documents with file search instead of the more expensive web search.
Concurrency is another limit. The same breakdown says session limits scale with lifetime spend: 10 concurrent sessions at no spend, 20 after $50, and 50 after $250. A launch-day surge can hit those ceilings if you have not pre-funded the account.
How Does the Price Compare With Other Voice APIs?
Grok Voice 3.0 sits in the middle of the pack once reasoning and extras are counted. The listed prices in the eesel breakdown include GPT-Live-1 at $0.05 per minute, ElevenLabs Agents at $0.08, and Deepgram Voice Agent at $0.075, though each bills the language model and phone line differently.
Artificial Analysis offers a cleaner cost comparison per hour of input audio, excluding tool calls. Grok Voice 3.0 comes to $4.80. GPT-Live-1 at low reasoning is $4.47, and at a higher setting it is $5.83. Gemini 3.8 Live at high thinking is $3.50, and GPT-Realtime-2.1 at high reasoning is $10.75.
Price is only one input. A cheaper model that resolves fewer calls can cost more overall through longer calls and human handoffs. Our review of whether AI voice agents are ready for production phone calls covers the resolution question in more depth.
What About the Company-Reported Results?
The deployment claims are impressive, and they are all self-reported. xAI says Starlink support handles about 20,000 customer calls per day, with roughly 70% resolved without a human. It also says about one in four sales calls ends in a purchase, and that Tesla's in-car assistant handles millions of daily conversations.
No independent audit of those figures has been published, according to the AI Weekly alert. The same alert carries an editor's note saying there is no model card or security audit, and that such a system can issue refunds and ship hardware. It calls that a governance gap.
That concern is worth taking seriously for any team giving a voice agent real authority. Limit what the agent can do without a human, log every action, and test refusal behavior before launch.
Should You Switch to Grok Voice 3.0?
Consider it if you already run Grok Voice and want better task scores at the same price. Hold off if low latency is your top priority, or if you need audited safety documentation.
A practical test plan runs about a week:
- Replay a sample of real call transcripts or recordings through both versions.
- Measure resolution rate, handoff rate, and average call length.
- Calculate cost per resolved call, including carrier and tool charges.
- Check latency from your own region, not just the vendor's number.
- Pin your model version if the platform allows it.
At Model Drop, our editors will revisit Grok Voice 3.0 as independent data arrives. Your next step is to calculate your cost per resolved call, not your cost per minute.
