xAI · now SpaceXSI · Speech-to-speech
Grok Voice 3.0 API Pricing & Cost Calculator
Grok Voice 3.0 (grok-voice-3.0) is xAI's speech-to-speech model for real-time voice agents, released Oct 14, 2026.
On the Voice Agent API it costs $0.08 per minute ($4.80 per hour) plus $0.004 per text input — the same price as Grok Voice 2.0.
Rate ruler · one hour of call time
$4.80
0:00 · Rate card
Grok Voice 3.0 API pricing per minute
Voice Agent API · list price
$0.08 / min
= $4.80 / hour
+ $0.004 per text input
- Model id
- grok-voice-3.0
- Alias
- grok-voice-latest → grok-voice-3.0
- Region
- us-east-1
- Released
- Oct 14, 2026
- Grok Voice 2.0
- Deprecated · $0.08 / min · same price
Billed per use — not in the calculator
- Web searchNot in calculator$5 / 1,000 calls
- X searchNot in calculator$5 / 1,000 posts · $10 / 1,000 profiles
- File search (collections)Not in calculator$2.50 / 1,000 calls
- Remote MCP toolsNot in calculatorNo call fee · tokens billed
Other xAI speech APIs
- Speech to Text$0.10 / hr (REST) · $0.20 / hr (streaming)
- Text to Speech$15.00 / 1M characters
1:00 · Call cost
What will your voice agent cost?
Conversations your agent handles in a day.
Minutes per call, one decimal allowed: 2.5 is 2 min 30 s.
Text you send into the call, such as context updates. Often 0.
Days in the period you're budgeting.
Comparison assumption
Share of each call the agent speaks. Only Gemini bills output audio separately.
Statement · Grok Voice 3.0 · Voice Agent API
1,000 calls/day × 3 min × 30 days = 90,000 min
- Call time90,000 × $0.08 / min$7,200.00
- Text inputs0 × $0.004 / input$0.00
- Per call
- $0.2400
- Per day
- $240.00
- Period · 30 days
- $7,200.00
- Not included: Tool calls during a call (web search, X search, file search)
- xAI doesn't say how partial minutes are billed; we pro-rate them.
Same calls elsewhere
These are not complete bills. Each row prices only what the vendor publishes per minute; the charges listed under it are metered another way and left out.
- OpenAIGPT-Live-1$4,500.00
List price only
- + The backend model's tokens, at that model's rates
- + Tool calls
- GoogleGemini 3.8 Live$1,260.00
List price only · assumes 50% agent talk
- + Text and thinking tokens ($0.75 in · $4.50 out per 1M)
- + Google Search grounding beyond 5,000 free requests a month ($14 / 1,000)
Running a voice agent in production?
2:00 · Billing models
How each vendor bills a voice minute
xAI
Grok Voice 3.0
One meter runs for the whole call
- Call time$0.08 / min
- Text input$0.004 each
Not in the per-minute rate:
- Tool calls during a call (web search, X search, file search)
xAI doesn't say how partial minutes are billed; we pro-rate them.
OpenAI
GPT-Live-1
A session meter, plus the model behind it
- Session time$0.05 / min
- Backend modelPer token, billed separately
Not in the per-minute rate:
- The backend model's tokens, at that model's rates
- Tool calls
Billed per second (OpenAI).
Gemini 3.8 Live
Listening and speaking are metered apart
- Audio in$0.005 / min
- Audio out$0.018 / min
- Text and thinkingPer token, billed separately
Not in the per-minute rate:
- Text and thinking tokens ($0.75 in · $4.50 out per 1M)
- Google Search grounding beyond 5,000 free requests a month ($14 / 1,000)
Google bills audio per token and lists these per-minute equivalents.
Paid-tier prices; Google also lists a free tier.
Extended Thinking's background reasoning is billed as thinking tokens.
Sticker hour vs measured hour
The sticker hour is what the published per-minute rates add up to over 60 minutes. The measured hour is what Artificial Analysis recorded per hour of input audio in its benchmark runs. The gap is what the per-minute rates leave out, such as backend-model and thinking tokens.
Measured hours: Artificial Analysis: Speech to Speech leaderboard ↗, accessed Oct 16, 2026.
3:00 · Quality vs cost
Grok Voice 3.0 vs GPT-Live-1 vs Gemini 3.8 Live
Cheap per minute and cheap per conversation are different claims. The plot puts each configuration’s measured hour against its score on the Speech to Speech Index; up and to the left is better value.
- 1Grok Voice 3.086.9 · $4.80 / hr
- 2Grok Voice 2.081.3 · $4.80 / hr
- 3GPT-Live-1 · Astra, medium83.2 · $5.83 / hr
- 4GPT-Live-1 · Sol, low80.1 · $4.47 / hr
- 5Gemini 3.8 Live Extended Thinking · High82.6 · $3.50 / hr
- 6Gemini 3.8 Live76.0 · $0.84 / hr
| Configuration | Speech to Speech Index | Speech reasoning | Conversational dynamics | Agentic performance | Task success rate | Arena preference (Elo) | Time to first audio | Cost per hour of input audio |
|---|---|---|---|---|---|---|---|---|
| Grok Voice 3.0 | 86.9 | 99% | 97.0% | 71.6% | 94.8% | 1070 | 0.81 s | $4.80 |
| Grok Voice 2.0 | 81.3 | 97% | 95.1% | 56.5% | 94.6% | 1011 | 0.70 s | $4.80 |
| GPT-Live-1Astra, medium | 83.2 | 90% | 94.9% | 74.5% | 87.4% | 1056 | 1.34 s | $5.83 |
| GPT-Live-1Sol, low | 80.1 | 89% | 97.3% | 59.3% | 90.9% | 1057 | 1.24 s | $4.47 |
| Gemini 3.8 Live Extended ThinkingHigh | 82.6 | 98% | 91.9% | 68.6% | 89.1% | 993 | 1.35 s | $3.50 |
| Gemini 3.8 Live | 76.0 | 92% | 96.1% | 30.1% | 93.2% | 1088 | 1.18 s | $0.84 |
4:00 · Use the API
How to use the Grok Voice 3.0 speech-to-speech API
Hear it first
Open the voice agent Playground in xAI's console to try the voices and your prompt before you write code.
Connect with a pinned model
Open a WebSocket to wss://api.x.ai/v1/realtime?model=grok-voice-3.0 with your API key. grok-voice-latest points to 3.0 today but moves with every release, so pin the version in production.
Configure the session
Send session.update with a voice, short instructions and server_vad turn detection, which handles barge-in. Reasoning is on by default at high effort; set reasoning.effort to none for faster turns.
Keep your key on the server
Browsers can't set WebSocket headers. Mint a short-lived ephemeral token on your server and pass it as the WebSocket protocol, prefixed xai-client-secret.
Quick start
import WebSocket from "ws"
const ws = new WebSocket("wss://api.x.ai/v1/realtime?model=grok-voice-3.0", {
headers: { Authorization: `Bearer ${process.env.XAI_API_KEY}` },
})
ws.on("open", () => {
ws.send(JSON.stringify({
type: "session.update",
session: {
voice: "eve",
instructions: "You are a concise support agent.",
turn_detection: { type: "server_vad" },
reasoning: { effort: "high" },
},
}))
})
ws.on("message", (raw) => {
const event = JSON.parse(raw)
if (event.type === "response.output_audio.delta") {
// play the audio chunk as it arrives
}
})Sockets and specs
- Model
- grok-voice-3.0 · alias grok-voice-latest
- Endpoint
- wss://api.x.ai/v1/realtime
- Region
- us-east-1
- Connect over
- WebSocket · SIP · LiveKit
- Output audio
- PCM, 8–48 kHz · G.711 μ-law · G.711 A-law · Opus
- Tools in a call
- web_search · x_search · file_search · mcp · function
Coming from OpenAI Realtime? Change the base URL and keep your event handling, then check xAI's notes on renamed and unsupported events. xAI also suggests shortening system prompts ported from GPT models: 3.0 needs fewer workarounds.
5:00 · Versions
Grok Voice versions: what changed in 3.0
Between model versions
- 97 days from Grok Voice 1.0 to Grok Voice 2.0
- 77 days from Grok Voice 2.0 to Grok Voice 3.0
| Metric | Grok Voice 2.0 | Grok Voice 3.0 | Change | Direction |
|---|---|---|---|---|
| Speech to Speech Index | 81.3 | 86.9 | +5.6 | Better |
| Speech reasoning | 97% | 99% | +2 pts | Better |
| Conversational dynamics | 95.1% | 97.0% | +1.9 pts | Better |
| Agentic performance | 56.5% | 71.6% | +15.1 pts | Better |
| Task success rate | 94.6% | 94.8% | +0.2 pts | Better |
| Arena preference (Elo) | 1011 | 1070 | +59 | Better |
| Time to first audio | 0.70 s | 0.81 s | +0.11 s | Worse |
| Cost per hour of input audio | $4.80 | $4.80 | ±$0.00 | Same |
6:00 · Quick answers
Grok Voice 3.0 FAQ
- How much does Grok Voice 3.0 cost?
- $0.08 per minute ($4.80 per hour) on xAI's Voice Agent API, plus $0.004 per text input you send into a call. Tool calls such as web search are priced separately on xAI's pricing page. Grok Voice 2.0 cost the same.
- Is the Grok Voice API free?
- xAI's pricing page lists no free tier for speech-to-speech: calls are billed at $0.08 per minute. You can try the voices in the voice agent Playground in xAI's console before you build.
- What is the Grok Voice 3.0 model ID?
- grok-voice-3.0. The alias grok-voice-latest has pointed to it since Oct 14, 2026, so agents using the alias switched automatically. xAI recommends pinning the versioned name in production, because the alias moves with each release.
- Is Grok Voice 3.0 compatible with the OpenAI Realtime API?
- Yes. xAI's docs say most OpenAI Realtime client libraries work after you change the base URL to wss://api.x.ai/v1/realtime. Some event names differ and a few events aren't supported; xAI lists them in its compatibility notes.
- Which languages and voices does Grok Voice support?
- xAI's speech-to-speech docs say 20+ languages; its Voice API page says 25+. Built-in voices include Eve, Ara, Leo, Rex and Sal, the same roster as xAI's Text to Speech API, and you can clone a custom voice from about two minutes of speech.
- Can I try Grok Voice 3.0 without writing code?
- Yes. xAI's console has a voice agent Playground, and Grok Voice 3.0 also powers the voice mode in Grok Bot.
- Is Grok Voice 3.0 the same as Grok 3 voice mode?
- No. Grok 3 is an xAI chat model from 2025, and "Grok 3 voice mode" means talking to it in the Grok app. Grok Voice 3.0 is the speech-to-speech model xAI released on Oct 14, 2026 for building voice agents through its API.
— · Sources
Sources and open questions
Pages we read
- xAI: Introducing Grok Voice 3.0 (Oct 14, 2026)x.ai
- xAI: API pricing (Voice API)docs.x.ai
- xAI: Speech to Speech docsdocs.x.ai
- xAI: Grok Voice 3.0 model pagedocs.x.ai
- xAI: Grok Voice API overviewx.ai
- OpenAI: GPT-Live 1 model pagedevelopers.openai.com
- Google: Gemini Developer API pricingai.google.dev
- Artificial Analysis: Speech to Speech leaderboardartificialanalysis.ai
What xAI’s pages don’t say
- What counts as a billed minute?
- xAI lists $0.08 per minute but doesn't say whether that's connection time or audio streamed, or how partial minutes round. The calculator uses the call length you enter and pro-rates partial minutes.
- What counts as one text input?
- xAI charges $0.004 per text input without defining it further. We count each text message you send into a call.
- Are tool calls in a voice session billed extra?
- xAI prices web search, X search and file search per call on its API, but doesn't say whether voice sessions bill them the same way. The calculator leaves them out.
- How many languages?
- xAI's speech-to-speech docs say 20+; its Voice API page says 25+.
Independent site. Not affiliated with xAI (SpaceXSI), OpenAI or Google.