xAI released Grok Voice 3.0 on October 14, and the headline is easy to repeat: the price did not move. It still costs $0.08 per minute, or $4.80 per hour, the same as version 2.0, while the benchmark scores went up. Whether that adds up to a better phone agent for your business depends on details the launch coverage mostly skips.
What Grok Voice 3.0 Is
Grok Voice 3.0 is a speech-to-speech model sold through xAI’s Voice Agent API. It listens and answers in audio directly, without chaining a separate transcription step, a text model and a text-to-speech step. It also supports tool calling, so an agent can look up an order or trigger an action mid-conversation. The LLM Gateway listing still labels it beta.
One practical detail: agents set to the grok-voice-latest alias were moved to 3.0 automatically. If you run a Grok voice agent, you did not choose this upgrade, so re-test your prompts and call handling.
The Grok Voice 3.0 Numbers, Side by Side
Per the breakdown in we0.ai’s pricing analysis, here is how 3.0 compares with 2.0:
- Artificial Analysis Speech-to-Speech Index: 86.9, up from 81.3
- Big Bench Audio (speech reasoning): 98.5%, up from 97.2%
- Full Duplex Bench (handling interruptions): 97.0%, up from 95.1%
- tau-Voice (customer-support tasks): 71.6%, up from 56.5%
- Time to first audio: 0.81 seconds, slightly slower than 2.0’s 0.70
The biggest jump is on tau-Voice, the support-task test, which is the one most relevant to businesses. The latency line cuts the other way. xAI says responses come in under a second and claims about 40% lower latency than OpenAI’s GPT-Live-1, but its own predecessor was a little faster on that measure.
Where It Trails OpenAI
On tau-Voice, Grok Voice 3.0 scores 71.6% against 74.5% for GPT-Live-1 (Astra, medium), so OpenAI still leads the support benchmark. Where Grok wins is cost. AI Weekly reports an average of $0.025 per Big Bench Audio task, against $0.030 for GPT-Live-1. The we0.ai breakdown puts GPT-Realtime-2.1 High at $0.055 for the same task.
Benchmark definitions differ, so these percentages are not interchangeable. A model that wins a speech-reasoning test can still stumble on a real call with a mumbling customer and a bad line.
What the $0.08 a Minute Does Not Cover
The per-minute rate covers the model. It does not cover telephony, infrastructure, external tools or human escalation. The pricing page also lists a separate $0.004 charge per text input in voice sessions. A realistic budget for a phone agent includes the phone carrier, your own backend and a person who takes over when the agent gets stuck.
For a quick estimate, take your average call length and multiply by $0.08. A five-minute call costs 40 cents in model fees before anything else is added.
Treat the Deployment Claims Carefully
xAI cites some large numbers: millions of daily conversations in Tesla vehicles, and about 20,000 Starlink support calls a day with roughly 70% resolved without a human. It also says one in four sales calls ends in a purchase. All of this is company-reported, and none of it has been independently audited.
Starlink is a single deployment with xAI’s own engineering support. It does not show that a small business with a messy knowledge base will get the same resolution rate. AI Weekly also notes that no model card or security audit has been published, while the system can already issue refunds and ship hardware. That is a governance question worth asking before you give any voice agent real permissions, a theme that also runs through our look at always-on agents and their limits.
How to Test a Voice Agent Before Trusting It
- Write down the ten most common reasons customers call you.
- Run each as a scripted call, including one with background noise and one where the caller interrupts.
- Check which actions the agent can take on its own, and limit refunds or account changes to a hard cap.
- Set a clear handoff rule so a human takes over after two failed attempts.
- Re-run the same calls after any model update, since an alias like
grok-voice-latestcan change underneath you.
The Bottom Line
Grok Voice 3.0 is a real improvement on paper at an unchanged price, and the support-task score is the number to watch. It is still second to OpenAI on that benchmark, and the strongest evidence for it comes from xAI itself. If you are comparing voice agents, treat it as a strong, cheap candidate to test, not a proven winner. For broader context on where this category came from, see our earlier piece on voice AI assistants, and for how another lab handles agents that act on their own, read Google’s Gemini Agent. Memory is the other weak point for any phone agent that must recall past calls, covered in why agents keep forgetting things.
Sharing Is Caring