Grok Voice 3.0 for Voice Agents: Upgrade Notes and Pricing

xAI, which now appears as SpaceXSI in its API documentation, has released Grok Voice 3.0, the third generation of its real-time speech-to-speech model. It went live on October 14 in the Voice Agent API, priced the same as the model it replaces. For teams building phone support lines, in-app assistants or sales agents, it is a drop-in upgrade. Independent testing from Artificial Analysis shows a clear jump in conversational quality.

• Model ID: grok-voice-3.0, with the grok-voice-latest alias now pointing to it automatically
• Price: $0.08 per minute of audio ($4.80 per hour), unchanged from Grok Voice 2.0
• Score: 86.9 on the Artificial Analysis Speech-to-Speech Index, up from 81.3, according to AI/TLDR and DevX
• Features: 20 languages with automatic detection, 28 built-in voices, and voice cloning from a clip up to 120 seconds long

What Is Grok Voice 3.0?

Grok Voice 3.0 is a speech-to-speech model. It listens to a caller and answers out loud in real time, with no separate transcription and text-to-speech steps. SpaceXSI's release notes say the API is unchanged. Any session using the grok-voice-latest alias was upgraded automatically on launch day. Grok Voice 2.0 is now deprecated, so agents pinned to that model ID should switch. The price stays the same either way.

Connections work over WebSocket, WebRTC and SIP. SIP lets the model plug straight into phone systems and contact-center platforms. DevX reports that server-side voice activity detection handles interruptions, so callers can talk over the agent the way they would with a person.

How Does Grok Voice 3.0 Compare on Benchmarks?

Artificial Analysis figures, as reported by AI/TLDR and DevX, show where the new model stands:

• Speech-to-Speech Index: Grok Voice 3.0 scores 86.9, ahead of GPT-Live-1 at 83.2 and Gemini 3.8 Live at 82.6
• Full Duplex Bench (turn-taking and interruptions): 97.0%
• Big Bench Audio (spoken reasoning): 98.5%
• Cost per Big Bench Audio task: $0.025, versus $0.030 for GPT-Live-1

Two numbers belong in any honest evaluation. On the τ-Voice support-call benchmark, Grok Voice 3.0 scores 71.6%, behind GPT-Live-1's 74.5%. Time to first audio also rose slightly, from 0.70 seconds to 0.81 seconds. DevX notes that this is still quicker than GPT-Live-1's 1.34 seconds. In short, Grok Voice 3.0 leads on conversation quality and price. Teams with complex multi-step support flows should run their own task tests before switching.

Multilingual Voices and Cloning

Language handling is one of the most practical upgrades. Grok Voice 3.0 supports 20 languages and detects them automatically. It can switch mid-call when a caller changes language, with no manual flag. Developers can pick from 28 built-in voices or clone a custom brand voice from a reference recording of up to 120 seconds.

SpaceXSI says its voice agent for Starlink handles about 20,000 calls a day and resolves roughly 70% without a human, according to AI/TLDR. That is a useful real-world signal of what production voice AI can already do.

Why Voice Agents Keep Getting Better

Voice is fast becoming the most competitive front in conversational AI. We have tracked GPT-Live-1's voice agent pricing and Gemini 3.8 Live this autumn. Each release has pushed quality up while holding prices steady or lowering them. Grok Voice 3.0 continues that pattern: a better model for the same $0.08 a minute. Follow the rest of the race in our AI news coverage.

Sources: SpaceXSI (xAI) API Release Notes — October 14, 2026; AI/TLDR — October 14, 2026; DevX — October 16, 2026.