Grok Voice 3.0
xAI's flagship speech-to-speech voice agent model — top of the AA speech index at unchanged $0.08/min pricing.
About
Grok Voice 3.0, announced October 14, 2026 by xAI (operating its AI unit as SpaceXSI/SpaceXAI), is a real-time speech-to-speech model for voice agents: it listens, reasons, calls tools, and speaks over a WebSocket connection with server-side voice activity detection and barge-in handling.
It is exposed as grok-voice-3.0 on the /v1/realtime endpoint, and the grok-voice-latest alias now points to it, so existing integrations upgraded without code changes. Reasoning runs at high effort by default with a none option for speed; tool use, web search, X search, remote MCP servers, and custom functions run inside the session. The 28 built-in voice IDs and custom voices from a 120-second reference clip carry over, and client apps can use short-lived ephemeral tokens.
xAI reports Grok Voice 2.0 is deprecated in favor of 3.0, with production use across Tesla vehicles and Starlink call centers: millions of daily conversations, roughly 70% of 20,000 daily support calls resolved without a human.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Independent (Artificial Analysis, Oct 13, 2026): #1 on the Speech-to-Speech Index at 86.9 vs GPT-Live-1 (Astra, Medium) 83.2 and Gemini 3.8 Live 82.6. Provider-reported: 97% on Full Duplex Bench; 71.6% on tau-Voice vs GPT-Live-1 at 74.5%; 74% on ServiceNow EVA-Bench.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Release facts (October 14, 2026; xAI; $0.08/min unchanged; grok-voice-latest auto-migration; Voice 2.0 deprecated; AA Speech-to-Speech 86.9) corroborated across Unite.AI, NeoTeo, and independent writeups, with the AA leaderboard itself independent. Docked for: (1) production usage figures (Tesla/Starlink volumes) are xAI's own claims, unverified; (2) no parameter counts or architecture details published; (3) strictly hosted, no weights.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 10.0/10
- Openness · license + docs
- 2.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Adoption 7: production use across Tesla and Starlink (millions of daily conversations per xAI); grok-voice-latest auto-migration
- · Maintenance 10: released 2026-10-14; current flagship voice model, 2.0 deprecated
- · Openness 2: hosted API only, no weights, no architecture details
- · Manual 8: release facts (Oct 14 2026; $0.08/min unchanged; AA Speech-to-Speech 86.9, independently leaderboarded) corroborated across Unite.AI, NeoTeo; docked for unverified xAI production figures, no architecture disclosure, closed weights