Grok Voice 3.0, the new speech-to-speech model from SpaceXSI, costs $0.08 a minute in the company’s Voice Agent API, or $4.80 an hour. That’s exactly what Grok Voice 2.0 cost, and developers whose agents call grok-voice-latest were moved to the new model on launch day, October 14, without changing a line of code.
The price isn’t the headline SpaceXSI wants. Its launch post calls Grok Voice 3.0 “the world’s most intelligent speech-to-speech model,” leans on a first-place ranking from the independent testing firm Artificial Analysis, and says the model already resolves most Starlink support calls without a human. Voice agents are where AI companies are going after call-center work most directly, so those claims will get quoted. Some hold up on the leaderboard itself. Some need an asterisk.
SpaceXSI is the company formerly known as xAI. According to SiliconANGLE, SpaceX acquired xAI in February, renamed it SpaceXAI in July and switched to SpaceXSI this month. Its developer docs still say “xAI API.”
Grok Voice 3.0 price: the same $4.80 an hour, and 2.0 is now deprecated
SpaceXSI’s pricing page, checked October 15, lists both voice models at the same rate:
| Model | Speech-to-speech price | Text input | Status |
|---|---|---|---|
grok-voice-3.0 |
$0.08 / min ($4.80 / hr) | $0.004 / text input | Current; grok-voice-latest points here |
grok-voice-2.0 |
$0.08 / min ($4.80 / hr) | $0.004 / text input | Deprecated |
The release notes say the API is unchanged and that upgrading means setting model to grok-voice-3.0 or the alias. They mark 2.0 as deprecated but, unlike the transcription model retired earlier this month, give no end-of-life date. The model page lists three regions: us-east-1, eu-west-1 and us-saltlake-2.
Cheaper than OpenAI, by Artificial Analysis’ meter. Not cheaper than Google
SpaceXSI says the model “costs less than other frontier models.” Per-minute and per-token prices don’t line up neatly, so the fairest yardstick is Artificial Analysis’ normalized figure: what each model cost to work through the same Big Bench Audio questions, scaled to an hour of input audio. From its leaderboard, checked October 15:
| Model (Artificial Analysis’ runs) | Cost per hour of input audio | Time to first audio |
|---|---|---|
| Grok Voice 3.0 | $4.80 | 0.81 s |
| Grok Voice 2.0 | $4.80 | 0.70 s |
| OpenAI GPT-Live-1 (Astra, medium) | $5.83 | 1.34 s |
| OpenAI GPT-Realtime-2.1 High | $10.75 | 1.21 s |
| Google Gemini 3.8 Live Extended Thinking (High) | $3.50 | 1.35 s |
The claim holds against OpenAI, including GPT-Live-1 in its GPT-6 Astra configuration. Google’s Gemini 3.8 Live with extended thinking comes in cheaper on this measure. The speed claim checks out: 0.81 seconds is about 40% less wait than GPT-Live-1’s 1.34 seconds, by our arithmetic, in line with SpaceXSI’s “40% lower latency” line. The catch is the row underneath. Artificial Analysis measured Grok Voice 2.0 at 0.70 seconds, so the new model is about a tenth of a second slower to start talking than the one it replaces.
First on the index, not first on every test
SpaceXSI’s launch table cites Artificial Analysis’ results dated October 13:
| Benchmark (SpaceXSI’s table, Artificial Analysis data) | Grok Voice 3.0 | Grok Voice 2.0 | GPT-Live-1 (Astra, Medium) | Gemini 3.8 Live (Ext. Thinking, High) |
|---|---|---|---|---|
| Speech-to-Speech Index | 86.9 | 81.3 | 83.2 | 82.6 |
| Big Bench Audio (reasoning) | 98.5% | 97.2% | 90.1% | 97.7% |
| Speech Agent Arena Elo | 1070 | 1011 | 1048 | 990 |
| Task success rate | 94.8% | 94.6% | 87.4% | 89.1% |
| Full Duplex Bench | 97.0% | 95.1% | 94.9% | 91.9% |
On Artificial Analysis’ own page, Grok Voice 3.0 does have the highest index score and the highest task success rate. It doesn’t lead everything. StepFun’s StepAudio 3 Realtime and two Alibaba Qwen models score higher on speech reasoning, and three models beat it on conversational dynamics. Two Gemini Live models rank above it on arena preference Elo, which the four-model table leaves out. The live page also shows slightly different Elo numbers for the two rivals (1056 and 993); Artificial Analysis says live Elo keeps moving, while the rating used in the index is frozen when a model first qualifies.
The closest race is τ-Voice, a test of simulated support calls with accents, noise and interruptions. SpaceXSI calls its 71.6% “comparable” with GPT-Live-1, but its own chart has OpenAI’s model ahead at 74.5%, and Artificial Analysis notes that figure comes from a single trial. On ServiceNow’s EVA-Bench for enterprise voice agents, SpaceXSI says Grok Voice 3.0 had the top accuracy score, 74%. Nitin Aggarwal, ServiceNow’s senior director of product management for agentic evals and RAI, is quoted in the post: “It pairs top accuracy with a natural, responsive conversation, setting a new standard for what enterprises can expect from AI on the front line of customer service.”
The Starlink numbers come from the parent company
The line The Verge picked out: Starlink call centers use Grok Voice to answer 20,000 calls a day across multiple languages, roughly 70% are resolved “with no human in the loop,” and one in four sales calls ends in a purchase, SpaceXSI says. A single agent built on 3.0 verifies customers, troubleshoots hardware, issues credits and ships replacement equipment. SpaceXSI also says Grok Voice handles millions of conversations a day in Tesla vehicles.
These are a vendor’s figures about its own parent’s call centers: Starlink is SpaceX’s satellite internet business, and SpaceX owns SpaceXSI. The post doesn’t define “resolved,” say what period the 70% covers, or say whether those calls ran on 3.0 or an earlier Grok Voice.
The next things to watch are a retirement date for grok-voice-2.0 and Artificial Analysis’ fuller τ-Voice runs for GPT-Live-1, which so far rest on a single trial. That’s the one test in SpaceXSI’s own charts where Grok still trails.
Frequently asked questions
How much does Grok Voice 3.0 cost?
$0.08 per minute, or $4.80 per hour, for speech-to-speech in SpaceXSI’s Voice Agent API. That’s the same as Grok Voice 2.0. The pricing page also lists $0.004 per text input.
How do I switch to Grok Voice 3.0?
Sessions that use the grok-voice-latest alias were upgraded automatically. Otherwise, set model to grok-voice-3.0. SpaceXSI says the API is unchanged and the price is the same.
Is Grok Voice 3.0 better than OpenAI’s GPT-Live-1?
On Artificial Analysis’ overall Speech-to-Speech Index, yes: 86.9 against 83.2 for GPT-Live-1 (Astra, medium). GPT-Live-1 scores higher on τ-Voice, 74.5% to 71.6%, but takes 1.34 seconds to start answering against Grok’s 0.81 seconds.
Is Grok Voice 3.0 faster than Grok Voice 2.0?
Not by Artificial Analysis’ measurement. It recorded 0.81 seconds to first audio for 3.0 and 0.70 seconds for 2.0. Version 3.0 scores higher on every part of the index, and on τ-Voice it jumps from 56.5% to 71.6%.
Where is Grok Voice 3.0 available?
In SpaceXSI’s Voice Agent API and, per the company’s launch thread on X, its Voice Agent Builder. The model page lists the us-east-1, eu-west-1 and us-saltlake-2 regions.
Sources
All accessed October 15, 2026, between about 8:45 and 8:55 AM ET.
- SpaceXSI, “Introducing Grok Voice 3.0” (Oct 14, 2026; price, benchmark table, τ-Voice chart, EVA-Bench, ServiceNow quote, Tesla and Starlink figures): x.ai/news/grok-voice-3
- SpaceXSI docs, Grok Voice 3.0 model page (model name, regions): docs.x.ai
- SpaceXSI docs, release notes, October 2026 (alias switch, unchanged API, 2.0 deprecation): docs.x.ai/developers/release-notes
- SpaceXSI docs, pricing (Voice Pricing table): docs.x.ai/developers/pricing
- Artificial Analysis, Speech to Speech leaderboard (index, components, time to first audio, cost per hour of input audio, trial notes): artificialanalysis.ai
- SpaceXSI on X, launch thread, Oct 14, 2026, 1:04 PM ET: launch post, latency and Starlink post, pricing and availability post
- SiliconANGLE, Duncan Riley, “Grok Voice 3.0 from SpaceXSI outscores OpenAI and Google on speech-to-speech index” (Oct 14, 2026; company name history): siliconangle.com
- The Verge, Jay Peters, “Starlink uses Grok’s voice model to handle customer service calls” (Oct 14, 2026): theverge.com
- Tesla on X, Grok with Connectors video (Sept 22, 2026; featured image source, linked from SpaceXSI’s post): x.com/Tesla
Related: GPT-6.1 Sol Ultrafast price vs Astra · Gemini 4 Argon price and benchmarks · More LLM news
Last updated: October 15, 2026. To report an error, see our corrections page. Articles are drafted with AI assistance and reviewed and edited by an editor; see our editorial policy.
