AI Industry
xAI Grok Voice 3.0 Tops Speech-to-Speech Index at $0.08/Min
Oct 15, 2026·4 min read·By Rahish Saifi
xAI released Grok Voice 3.0 on October 14, 2026, calling it the world’s most intelligent speech-to-speech model for real-time voice agents. The model targets sub-second latency, keeps the Voice Agent API at $0.08 per minute ($4.80 per hour)—the same rate as Grok Voice 2.0—and is already routing production traffic from Tesla vehicles and Starlink call centers, according to the company’s announcement. Benchmark figures below are xAI-cited scores from Artificial Analysis and ServiceNow unless noted.
Developers using the alias grok-voice-latest are on the new model without code changes, per xAI release notes; explicit grok-voice-2.0 slugs are deprecated in favor of grok-voice-3.0 or the latest alias at unchanged pricing.
What changed in the API
xAI said Grok Voice 3.0 is available now in the Voice Agent API with the same request shape as Grok Voice 2.0. Release notes state grok-voice-3.0 supports speech-to-speech at the same price as its predecessor, that grok-voice-latest now points to grok-voice-3.0, and that sessions on the alias upgrade automatically.
For teams still pinning grok-voice-2.0, xAI marks that slug deprecated and directs upgrades to grok-voice-3.0 or grok-voice-latest. The Grok Voice 3.0 model page documents the slug for new integrations.
Artificial Analysis speech-to-speech scores
In independent testing published October 13, 2026, Artificial Analysis ranks Grok Voice 3.0 #1 on its Speech-to-Speech Index at 86.9, ahead of Grok Voice 2.0 (81.3), GPT-Live-1 (Astra, Medium) (83.2), and Gemini 3.8 Live (Ext. Thinking, High) (82.6), xAI reported. The index blends speech reasoning, τ-Voice, and preference and task success from blind head-to-head calls.
xAI also highlighted a 97% score on Full Duplex Bench—a measure of pauses, turn-taking, and interruptions—and published this comparison table from Artificial Analysis:
| Benchmark | Grok Voice 3.0 | Grok Voice 2.0 | GPT-Live-1 (Astra, Medium) | Gemini 3.8 Live (Ext. Thinking, High) |
|---|---|---|---|---|
| Overall AA Speech-to-Speech Index | 86.9 | 81.3 | 83.2 | 82.6 |
| Speech Reasoning Big Bench Audio | 98.5% | 97.2% | 90.1% | 97.7% |
| Human Preference Speech Agent Arena Elo | 1070 | 1011 | 1048 | 990 |
| Task Completion Task Success Rate | 94.8% | 94.6% | 87.4% | 89.1% |
| Conversational Dynamics Full Duplex Bench | 97.0% | 95.1% | 94.9% | 91.9% |
τ-Voice scores how agents handle support calls with diverse accents, background noise, and interruptions. Grok Voice 3.0 reaches 71.6% on τ-Voice—comparable with GPT-Live-1 (Astra) in xAI’s summary—but at 40% lower latency, which the company frames as a better intelligence-versus-responsiveness tradeoff.
On cost, Artificial Analysis measures what each model spends to complete a task in its Big Bench Audio run. xAI said Grok Voice 3.0 comes in under both GPT-Live-1 (Astra, Medium) and GPT-Realtime-2.1 on that cost-per-task metric.
Enterprise voice-agent accuracy
On EVA-Bench, ServiceNow’s benchmark for enterprise voice agents, Grok Voice 3.0 posts the highest accuracy score at 74% and is the best-performing model when jointly scored across accuracy and customer experience, according to xAI.
“Grok Voice 3.0 sets a new Pareto frontier on EVA-Bench, ServiceNow’s open-source benchmark scoring voice agents on accuracy and experience across 200+ realistic enterprise scenarios. 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.”
— Nitin Aggarwal, Senior Director of Product Management, Agentic Evals and RAI, ServiceNow
Production volume at Tesla and Starlink
xAI said Grok Voice already powers millions of conversations a day in Tesla vehicles, where drivers can use Connectors for calendar or email checks and Grok Bot for larger tasks such as ordering coffee or booking dinner hands-free. Grok Voice 3.0 also backs Grok Bot’s voice mode.
At Starlink, call centers use Grok Voice to answer 20,000 customer calls every day across multiple languages. Roughly 70% are resolved with no human in the loop, xAI said, and one in four sales calls ends in a purchase before the caller hangs up. The company described a single agent built on Grok Voice 3.0 that verifies customers, troubleshoots hardware, issues credits, and ships replacement equipment.
Pricing and rollout
Grok Voice 3.0 is priced at $0.08 / min ($4.80 / hr) in the Voice Agent API—the same as Grok Voice 2.0. Agents on grok-voice-latest were switched automatically with no integration changes required, xAI said.
Latency is class-leading at under one second in the announcement’s wording—useful context for builders comparing streaming speech-to-speech stacks, though your end-to-end numbers will still depend on network, telephony, and tool round-trips.
What it means for developers and businesses
If you already pay for Grok Voice through the latest alias, you likely received a capability bump without a pricing event—the same $0.08/min rate as Grok Voice 2.0. Teams comparing vendors should still run their own evals: Artificial Analysis and EVA-Bench numbers are useful directional signals, but accents, compliance scripts, and CRM tools in your stack can diverge from public leaderboards.
For contact-center and field-service builders, the τ-Voice and Full Duplex Bench story matters as much as raw reasoning scores: interruptions, barge-in, and hold-music environments break naive pipelines. xAI’s Starlink and Tesla anecdotes are company-reported production metrics, not third-party audits, but they signal where xAI is stress-testing duplex behavior at scale.
Deprecation of grok-voice-2.0 is a nudge to pin grok-voice-3.0 in staging, re-run latency and cost traces, and confirm fallback behavior if you need reproducible model IDs for regulated workflows.
What to watch
Watch whether competitors respond on the Artificial Analysis index after the October 13, 2026 leaderboard date, and whether xAI publishes third-party EVA-Bench artifacts beyond the ServiceNow quote. If you operate voice agents in the EU or under sector rules, validate retention, logging, and human-handoff paths independently—the API surface did not change, but model behavior can.