Gemini 3.8 Live: Voice Models That Think While They Talk

Google released two voice models for real-time conversation. Gemini 3.8 Live Extended Thinking reasons while it speaks instead of going silent, and tops Artificial Analysis’ speech index at 82.6.
artificial-intelligence
Author

Kabui, Charles

Published

2026-09-19

Keywords

gemini-3-8-live, voice-agents, speech-to-speech, google-deepmind, real-time-ai

Download as PDF

Google released two new voice models on September 15: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. Both are built for spoken conversation that feels like a phone call rather than a chatbot prompt. 3.8 Live is the cheaper, faster model for high-volume use; Extended Thinking handles harder, multi-step jobs. The clever part is how it hides the work. Instead of going quiet while it thinks, it reasons and speaks at the same time, dropping cues like “Let me check that…” while background tool calls finish. Both auto-detect 97 languages mid-call. Google says Extended Thinking tops Artificial Analysis’ Speech to Speech Index at 82.6, ahead of OpenAI’s GPT-Live-1 at 81.5.

For anyone building a phone agent, the failure that matters is silence. An assistant that thinks for eight seconds sounds broken, and callers hang up. Reasoning in parallel with speech, plus tool calls that resolve in the background, removes that dead air without making the model dumber. Google is shipping both widely: the Gemini API and AI Studio for developers, Search Live and Gemini Live for everyone, and Workspace apps like Docs, Gmail, and Keep for AI subscribers.

The race is now about conversation, not just answers. Google holds the top two spots on the Speech Agent Arena, though the older, cheaper 3.1 Flash Live Minimal ties 3.8 Live on preference. On ServiceNow’s EVA-Bench, no system clears 0.6 on both accuracy and conversational quality. Every audio clip Google generates is SynthID watermarked.

Read More: Gemini 3.8 Flash: Google’s Third Flash in Six Weeks, With a Defender-Only Cyber Model

Sources:

Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service

Reuse

GNU GENERAL PUBLIC LICENSE v3.0(View License)

Citation

BibTeX citation:
@misc{kabui2026,
  author = {{Kabui, Charles}},
  title = {Gemini 3.8 {Live:} {Voice} {Models} {That} {Think} {While}
    {They} {Talk}},
  date = {2026-09-19},
  url = {https://toknow.ai/posts/gemini-3-8-live-extended-thinking-voice-agents/},
  langid = {en-GB}
}
For attribution, please cite this work as:
Kabui, Charles. 2026. “Gemini 3.8 Live: Voice Models That Think While They Talk.” https://toknow.ai/posts/gemini-3-8-live-extended-thinking-voice-agents/.