← Tools

Grok Voice Agent API

xAI

★★★★☆

xAI's realtime speech platform for tool-enabled voice agents, telephony, TTS, and STT.

Category audio
Pricing Voice Agent $0.05/minute plus $0.004 per realtime text message; TTS $15/1M characters; STT $0.10/hour REST or $0.20/hour streaming.
Status active
Platforms api
xai grok voice speech-to-speech tts stt realtime telephony sip websocket multilingual custom-voices tools
Updated July 10, 2026 Official site →

Overview

Freshness note: AI audio and realtime APIs evolve rapidly. This profile is a point-in-time snapshot last verified on July 10, 2026.

Grok Voice Agent API is xAI’s developer platform for realtime speech-to-speech agents, text-to-speech, speech-to-text, telephony, and custom voices. The core agent interface streams audio and text bidirectionally over WebSocket, while separate /v1/tts and /v1/stt routes cover synthesis and transcription without requiring a full conversational session.

The current recommended realtime model is grok-voice-think-fast-1.0; the grok-voice-latest alias follows the newest model, while the older grok-voice-fast-1.0 is deprecated. Teams that need reproducible production behavior should pin the versioned model rather than silently following the alias.

For a hosted configuration and operations layer, Grok Voice Agent Builder packages the same voice foundation with a no-code UI, phone numbers, knowledge collections, tools, guardrails, and call review.

Key Features

The realtime endpoint is wss://api.x.ai/v1/realtime. Server applications can authenticate with an API key, while browser and mobile clients should use short-lived ephemeral tokens so the main credential is never exposed. Session controls cover instructions, voice selection, reasoning effort, voice-activity detection, audio formats, pronunciation replacements, resumption, and tool configuration.

The agent can use file search, web search, X search, remote MCP tools, and custom functions during a conversation. Direct SIP supports phone-agent deployments, and xAI documents compatibility with much of the OpenAI Realtime API event model, with a published list of naming differences and unsupported events.

xAI’s July catalog update added 21 multilingual voices to the original five, for 26 built-in voices available across realtime, TTS, and Builder. Custom voices can be created from a reference clip up to 120 seconds long—xAI recommends 90–120 seconds—and used in realtime or TTS, subject to account limits and regional availability.

The broader Voice API also includes expressive TTS with speech tags and telephony codecs, plus batch and streaming STT with word timestamps, multichannel transcription, and formatting support.

Strengths

The platform’s strongest feature is the integrated speech-to-speech path. Tool use, reasoning, turn handling, and audio generation stay inside one realtime session instead of being stitched across separate vendors. The separate TTS and STT routes still let teams use only the component they need.

The transport and migration story is practical: WebSocket, SIP, ephemeral client tokens, and partial OpenAI Realtime compatibility reduce the amount of custom infrastructure required for a first production integration.

Limitations

Realtime voice remains operationally demanding. Latency, interruption behavior, silence thresholds, phone-quality audio, multilingual performance, retries, and tool failures all need end-to-end testing. Compatibility with OpenAI’s event model is useful but not exact, so migration should be verified event by event.

Custom voice cloning is currently documented as available only in the United States except Illinois, with a default limit of 30 custom voices per team, and it requires explicit consent and misuse controls. Voice and transcript data also need defined retention, access, and deletion policies. Pricing can scale quickly at call-center concurrency even when the per-minute rate looks simple.

Practical Tips

Start with grok-voice-latest during exploration, then pin grok-voice-think-fast-1.0 after evaluation. Use ephemeral tokens for client-side connections, keep permanent API keys server-side, and scope tools by task rather than exposing a broad business account.

Measure time to first audio, interruption recovery, task completion, transfer accuracy, tool error rate, and cost per resolved call. Test the exact codec and telephony path used in production; clean browser audio is not a substitute for a real phone line.

Verdict

Grok Voice Agent API is now a broad and credible speech platform rather than a single realtime endpoint. It is a strong fit for teams that need integrated voice reasoning, tools, SIP, and component APIs, provided they treat model pinning, permissions, consent, and call operations as first-class engineering work.