Grok Voice 3.0 Takes #1 on Speech-to-Speech Index at 86.9, Runs Millions of Tesla and Starlink Calls Daily
SpaceXSI (formerly xAI) released Grok Voice 3.0 on October 14 and rolled it to all agents using the grok-voice-latest alias without a price change. Artificial Analysis’s independent testing, conducted October 13, ranks it first on the Speech-to-Speech Index with a score of 86.9.
Benchmark Numbers
The AA Speech-to-Speech Index is an equal-weighted composite of Big Bench Audio speech reasoning, τ-Voice agentic performance, Speech Agent Arena human preference, and task success rate.
| Benchmark | Grok Voice 3.0 | Grok Voice 2.0 | GPT-Live-1 Astra | Gemini 3.8 Live |
|---|---|---|---|---|
| AA Speech-to-Speech Index | 86.9 | 81.3 | 83.2 | 82.6 |
| Big Bench Audio | 98.5% | 97.2% | 90.1% | 97.7% |
| τ-Voice | 71.6% | 56.5% | 74.5% | 68.6% |
| Full Duplex Bench | 97.0% | 95.1% | 94.9% | 91.9% |
| Task success rate | 94.8% | 94.6% | 87.4% | 89.1% |
| Arena Elo | 1070 | 1011 | 1048 | 990 |
| Time to first audio | 0.81s | 0.70s | 1.34s | 1.35s |
Source: Artificial Analysis, October 13, 2026.
The composite lead has an uneven edge. On τ-Voice — which benchmarks support call completion through accents, background noise and tool use — Voice 3.0 scores 71.6%, below GPT-Live-1 at 74.5%. Time to first audio improved versus the competition (GPT at 1.34s, Gemini at 1.35s) but regressed against Voice 2.0, which measured 0.70s. SpaceXSI attributes the τ-Voice gap to 40% lower latency on that test: the model reaches the correct database end state faster even if its task accuracy is second.
Production Scale
Tesla vehicles account for millions of Grok Voice conversations per day. Drivers use it for calendar queries and email. Larger tasks, including restaurant bookings, go through Grok Bot. Starlink handles around 20,000 customer calls per day on the model across multiple languages, with roughly 70% resolved without a human transfer. SpaceXSI also reports that one in four sales calls ends in a purchase.
These are vendor-reported figures from the launch announcement, without external audit.
On ServiceNow’s open-source EVA-Bench, which grades enterprise voice agents across more than 200 scenarios, Grok Voice 3.0 scored 74% accuracy. No model scored higher on the combined accuracy-and-customer-experience metric.
Pricing and API
Price holds at $0.08 per minute. For context, GPT-Live-1 lists at $0.05 per minute but charges separately for backend model and tool calls. Gemini 3.8 Live bills audio per minute plus text tokens. On Artificial Analysis’s Big Bench Audio cost-per-task measure, Grok Voice 3.0 runs at $0.025 per task versus $0.030 for GPT-Live-1 and $0.055 for GPT-Realtime-2.1.
The grok-voice-latest alias now points to 3.0. Grok Voice 2.0 is deprecated. Developers who pinned the versioned model string are unaffected; those on the alias alias switched automatically.
Company Context
SpaceXSI is the AI unit that emerged from xAI. The company was renamed SpaceXAI in July 2026 after SpaceX acquired the AI operation in February, and the SpaceXSI name came into use earlier in October. SpaceX also rebranded its broader AI effort under this umbrella, which includes Grok Bot and the Grok Developer Platform.
Grok Voice 3.0 is closed-weights and available only through the hosted Voice Agent API.