Grok Voice 3.0 kept 2.0's $0.08 a minute and took first place on Artificial Analysis. It is 18% cheaper per hour than GPT-Live-1 on Astra, and Gemini 3.8 Live with extended thinking scores 4.3 points lower for 27% less.

Same card, new model behind it

Prices from xAI's pricing page; scores from xAI's announcement table and chart citing Artificial Analysis on October 13; latency from the Artificial Analysis leaderboard. Read October 21, 2026.

The release note is one line: grok-voice-latest now points at grok-voice-3.0, same endpoint, same session events. If your code uses the alias, you've been on 3.0 for a week and your invoice didn't notice. That's the opposite of what happened on August 5, when the same kind of alias move took the rate from $0.05 to $0.08.

Two small costs come with the upgrade. Time to first audio went up by 0.11 seconds, which you'll hear on a phone line. And 2.0 is deprecated with no end date. Pinning it saves you nothing, since it bills the same $0.08. The only reason to stay on 2.0 is if 3.0 breaks a prompt you've tuned, and even then you're on borrowed time.

What an hour costs across the board

Per-minute prices don't compare cleanly. GPT-Live-1 lists at $0.05 a minute but bills the backend model on top, and Gemini Live and GPT-Realtime bill audio tokens. Artificial Analysis publishes one cost per hour of input audio for each, measured on its Big Bench Audio run, which captures all of that, so we used it. Here it is scaled to a support line doing 40,000 minutes a month, about 10,000 calls of four minutes each:

• Gemini 3.8 Live$560
• Gemini 3.8 Live Ext. Thinking (High)$2,333
• GPT-Live-1 (Sol, low)$2,977
• Grok Voice 3.0$3,200
• GPT-Live-1 (Astra, medium)$3,887
• GPT-Realtime-2.1 (High)$7,167

AA's hourly rates, in the same order: $0.84, $3.50, $4.47, $4.80, $5.83 and $10.75. Grok's figure is its list price; the others depend on how much the model talks.

Grok sits in the middle. It's $687 a month under GPT-Live-1 on Astra, $223 over GPT-Live-1 on Sol, and $867 over Gemini 3.8 Live with extended thinking. Plain Gemini 3.8 Live is in a different league on price at $560, but it scores 76.0, nearly 11 points behind.

Dollars per index point

Divide each hourly rate by the Speech to Speech Index score and the order changes a little:

Artificial Analysis Speech to Speech leaderboard, read October 21, 2026.

Grok and GPT-Live-1 on Sol are tied within a tenth of a cent, and Grok gets you 6.8 more points for 33 cents an hour. Against Astra it is 18% cheaper and 3.7 points better, which is the comparison xAI leads with. It's a fair one.

The model xAI's chart doesn't dwell on is Gemini 3.8 Live with extended thinking: 4.3 points lower, 27% cheaper. It is also slower, at 1.35 seconds to first audio against Grok's 0.81. On a phone call that half second matters more than four index points, in our view, so we'd still pick Grok for anything a customer talks to live.

Where the $0.08 isn't the whole bill

Reasoning is on by default at high effort, and there is no line for it. You can set reasoning.effort to "none" and the price doesn't move. xAI publishes no latency figure for that setting, so test it before you assume it buys back the 0.11 seconds.

Tools are separate. The voice API exposes web search, X search, file search, MCP and your own functions. xAI's general table lists web search at $5 per 1,000 calls and file search at $2.50, and nothing on the page exempts voice sessions. An agent that searches once per call adds half a cent to a 32-cent call. Your own function calls carry no xAI fee.

The $0.004 text fee is easy to miss on phone lines. xAI says SIP keypresses are "flushed to the model as text input". It batches presses until the caller hits #, pauses 2.5 seconds or starts talking, and doesn't say whether each flush is a billable text event. If it is, an IVR flow with five flushes per call adds 2 cents, or $200 a month on our 10,000 calls. Your carrier bills its own minutes too, since xAI doesn't provision numbers through the API.

We asked in August how xAI counts a minute, and the docs still don't say whether it is wall-clock time or audio sent plus audio received, or whether silence bills. We assume wall-clock here. If it's both streams, every Grok number above doubles.

Ten sessions until you spend $50

Speech to speech is rate limited by concurrent sessions alone. A new account gets 10. Tiers unlock on cumulative spend since January 1: 20 sessions at $50, 50 at $250, 100 at $1,000 and 200 at $5,000. Spend counts once it is paid, as prepaid credit or a settled invoice.

At $0.08 a minute, $250 is 3,125 minutes. A support line doing 40,000 minutes a month uses that up in under three days, so the limit only bites during a pilot. Plan the pilot for 10 concurrent callers, not your launch-day peak.

Pin, switch, or leave it alone

Getting 5.6 index points for nothing is a good release. If you already use grok-voice-latest, you don't need to do anything. If you pinned grok-voice-2.0, move to 3.0 on your own schedule rather than xAI's.

If you're choosing a voice stack today, Grok has the top score and the quickest first audio of the six we priced, at a mid-table price. Gemini 3.8 Live wins if cost matters more than the last few points or the last half second. Our calculator prices the token-billed models for your own call length.

Where the numbers come from

• xAI: Grok Voice 3.0 - October 14 launch, price, benchmark table citing Artificial Analysis
• xAI release notes and xAI pricing - alias change, 2.0 deprecation, voice and tool rates
• xAI speech to speech docs and rate limits - reasoning default, tools, SIP, concurrency tiers
• Artificial Analysis Speech to Speech leaderboard - index scores, time to first audio, cost per hour of input audio
• OpenAI pricing and Gemini API pricing - GPT-Live-1, GPT-Realtime-2.1 and Gemini 3.8 Live rates