xAI launches Grok Voice 3.0, topping Artificial Analysis speech index

xAI has released Grok Voice 3.0, a speech-to-speech model that ranks first on Artificial Analysis’s Speech-to-Speech Index with a score of 86.9. It responds in under one second and costs $0.08 per minute in the API, unchanged from Grok Voice 2.0.

What is new in Grok Voice 3.0

According to xAI, the model sets a record for resolving customer support calls and has class-leading latency. The company cites Artificial Analysis, an independent testing firm, whose index combines speech reasoning, the τ-Voice benchmark and preference and task success from blind head-to-head calls. In that testing, which xAI dates to October 13, 2026, Grok Voice 3.0 beat its predecessor (81.3) and the two rivals xAI lists: GPT-Live-1 in its Astra, Medium configuration (83.2) and Gemini 3.8 Live with extended thinking at a high setting (82.6).

The sub-scores in xAI’s table show where the gains are:

• Big Bench Audio (speech reasoning): 98.5%, against 97.2% for Grok Voice 2.0.
• Speech Agent Arena (human preference, Elo): 1070, against 1011 for the previous version and 1048 for GPT-Live-1.
• Task success rate: 94.8%, a small step up from 94.6%.
• Full Duplex Bench, which measures handling of pauses, turn-taking and interruptions: 97.0%, against 95.1%.

On τ-Voice, which tests support calls with varied accents, background noise and interruptions, Grok Voice 3.0 scores 71.6%. xAI says that is comparable with GPT-Live-1 (Astra) but at 40% lower latency.

xAI also points to cost. In Artificial Analysis’s Big Bench Audio run, which measures what each model spends to finish a task, Grok Voice 3.0 came in cheaper than both GPT-Live-1 (Astra, Medium) and GPT-Realtime-2.1, according to xAI.

ServiceNow’s open-source EVA-Bench, which scores enterprise voice agents across more than 200 scenarios, gave the model the highest accuracy at 74%. It also ranked best when accuracy and customer experience were scored together. ServiceNow’s Nitin Aggarwal said the model “pairs top accuracy with a natural, responsive conversation.”

Availability and deployments

The model is available now in the Voice Agent API at $0.08 per minute, or $4.80 per hour. Agents that use the grok-voice-latest alias were switched over automatically, so they need no changes.

xAI says Grok Voice already handles millions of conversations a day across Tesla and SpaceX. In Tesla vehicles, drivers can use Connectors for requests such as checking a calendar or answering an email, and Grok Bot for larger tasks such as ordering coffee or booking dinner, hands-free. The model also powers the voice mode of Grok Bot.

Starlink call centers use Grok Voice for 20,000 customer calls a day in multiple languages. According to xAI, about 70% of those calls are settled without a human involved. The company adds that 25% of sales calls result in a purchase while the customer is still on the line. In xAI’s description, one Grok Voice 3.0 agent covers identity checks, hardware troubleshooting, account credits and dispatching replacement gear.

Why it matters

For developers already on the Voice Agent API, the practical change is small. The grok-voice-latest alias moves them to the new model at the same $0.08 per minute. Teams that pin a version, or that compare vendors, face a harder job. OpenAI‘s developer documentation lists gpt-realtime audio at $32 per million input tokens and $64 per million output tokens, so a per-minute figure and a per-token figure cannot be compared without measuring real call lengths and how much each side talks. A pilot on actual call recordings will say more than any index score.

For businesses, the Starlink numbers are the closest thing to a reference deployment, but they have shifted. Crypto Briefing reported about 15,000 daily support calls, with about 70% of support issues handled without a human, and said callers began hearing the AI identify itself in mid-April 2026. The 20,000 figure in xAI’s launch material is higher, and both come from company statements without outside audit.

Companies that want AI agents on outbound sales calls face a legal constraint. A February 2024 FCC declaratory ruling, as summarized by law firm Steptoe, treats AI-generated voices as artificial voices under the Telephone Consumer Protection Act. Callers therefore generally need prior express consent, and state attorneys general can enforce the rule. Xai’s claim that 25% of sales calls convert on the line describes inbound calls. It does not show that the same approach works for outbound campaigns, which carry that consent burden.