# Grok Voice 3.0: A 0.81-Second Speech-to-Speech Model at $0.08 a Minute, and How to Read the Benchmarks

- Published: 16 October 2026
- Updated: 16 October 2026
- Author: RenderBob team
- Category: News
- Canonical: https://renderbob.ai/blog/grok-voice-3-0-speech-to-speech-pricing-benchmarks

xAI released Grok Voice 3.0, a speech-to-speech model for voice agents, at the same $0.08 a minute as its predecessor. It leads one independent index, and trails on other measures.

xAI released Grok Voice 3.0, a speech-to-speech model for voice agents, at the same price as its predecessor. SiliconANGLE dates the release to 14 October. The headline benchmark rank comes from Artificial Analysis, and the other numbers are mixed.

## What is reported

xAI's announcement prices Grok Voice 3.0 at $0.08 per minute through its Voice Agent API, or $4.80 per hour, unchanged from Grok Voice 2.0. It says latency is under one second. AlphaSignal, summarising Artificial Analysis, puts time to first audio at 0.81 seconds, slower than the 0.70 seconds it reports for version 2.0, and about 40% faster than the OpenAI and Google configurations in that comparison. xAI says that on its tau-Voice comparison the model has 40% lower latency than GPT-Live-1. AlphaSignal describes the API as compatible with OpenAI's Realtime API, and xAI says the grok-voice-latest alias now points to 3.0.

## The benchmarks, read carefully

xAI publishes this Artificial Analysis comparison, dated 13 October:

- Speech-to-Speech Index: 86.9, ranked first. GPT-Live-1 is listed at 83.2 and Gemini 3.8 Live at 82.6 in the same table.
- Big Bench Audio: 98.5%. Full Duplex Bench, which covers pauses, turn-taking, and interruptions: 97.0%.
- tau-Voice: 71.6%. xAI calls that comparable with GPT-Live-1. SiliconANGLE reports GPT-Live-1 at 74.5% on the same test.
- Task success: 94.8%. Preference Elo in xAI's table: 1070, ahead of the GPT-Live-1 and Gemini 3.8 Live rows shown there. AlphaSignal says a wider Speech Agent Arena preference ranking places it behind two Gemini live configurations.

The model leads one index and trails on other cuts of the same research. Which benchmark matters depends on the job: a support line cares about task success, an interactive character cares about how it sounds and feels.

## What is missing

The announcement does not mention voice cloning, custom voices, or consent requirements. The API marketing page lists 25 or more languages. The announcement itself does not set out regional availability. For any production use with a performer's voice, those are the questions that matter most. Our voices and transcript-layer piece sets out the checklist.

## Relevance to filmmakers

Speech-to-speech models are mostly built for agents, but they matter for interactive characters, installations, and live demos where a figure must answer in real time. The 0.81-second figure is a full speech-in, speech-out loop. The 150 ms time-to-first-speech that ElevenLabs reports for Eleven v4 Turbo is text-to-speech startup.

## What to watch

Independent listening tests, language coverage for the speech-to-speech model, a voice-consent policy, and whether xAI prints a release date on its own page. SiliconANGLE's date is 14 October.