Gemini 3.8 Live Extended Thinking Is the Voice-Agent Model to Start With
- What happened
- Google began rolling out Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking in the Gemini API and Google AI Studio on 15 September 2026.
- Why it matters
- Google says Extended Thinking ranks first on Artificial Analysis's Speech-to-Speech Quality Index (82.6) and reasons while it keeps talking.
- What to do
- Trial Extended Thinking first in AI Studio on your own voice flows, and keep 3.8 Live for latency-critical turns.
Our verdict: start with Gemini 3.8 Live Extended Thinking. We rate it Recommended: Google says it tops Artificial Analysis's Speech-to-Speech Quality Index, and that it reasons and speaks at the same time. Its sibling Gemini 3.8 Live is Conditional: pick it only when turn latency matters more than reasoning depth. If your voice stack runs on GPT-Live-1 or Grok Voice Think Fast 2.0, these are the two models to benchmark against.
What happened
On 15 September 2026, Google began rolling out two real-time voice models to the Gemini API and Google AI Studio.
| Gemini 3.8 Live | Gemini 3.8 Live Extended Thinking | |
|---|---|---|
| Google's headline claim | 2nd in Speech Agent Arena | 1st on the Artificial Analysis Speech-to-Speech Quality Index (82.6) |
| Other figures Google cites | 97 languages, detected and switched mid-conversation | τ-Voice 68.6%, Sierra τ-Voice-banking 35.1%, Big Bench Audio 97.7% |
Every figure above is as Google cites it. We have not run these benchmarks ourselves.
Google says Gemini 3.8 Live runs tool and API calls in the background without breaking the conversation, and that Extended Thinking reasons and speaks at the same time. Google also says all audio its AI products generate carries a SynthID watermark.
Where you get them
- Developers. Gemini API and Google AI Studio, rolling out from 15 September.
- Enterprise. Private preview in Gemini Enterprise. Related: Live Avatar for Gemini 3.8 Live in Gemini Enterprise.
- Consumers. Gemini 3.8 Live powers Search Live. Extended Thinking runs Gemini Live, Docs for Google AI Pro and Ultra subscribers, and Gmail and Keep for all Google AI subscribers.
Why it matters
Voice agents have forced a trade between talking fast and thinking hard. Extended Thinking is Google's bet that you no longer have to pick: Google says it reasons and speaks at the same time. A first place on a third-party speech-to-speech index backs that bet, though the ranking reaches us through Google's own post.
Google's launch post names no API model IDs and no rate. Check the model pages in our directory for current pricing and the Live API limits before you plan volume.
What changes for you
- Building a new voice agent on Google. Default to Extended Thinking and only drop to 3.8 Live if your latency budget fails.
- Already on GPT-Live-1 or Grok Voice Think Fast 2.0. Run your own call flows through both Gemini models in AI Studio before switching. Google's scores are not your traffic.
- Shipping consumer audio. Google says all audio its AI products generate carries a SynthID watermark, so plan for it on every clip.
What to do
- 1 Test Gemini 3.8 Live Extended Thinking in Google AI Studio on your own voice-agent flows. Google's benchmark figures are its own citations.
- 2 Use Gemini 3.8 Live only where turn latency matters more than reasoning depth.
- 3 Check current rates and Live API limits on the directory model pages before planning volume.
Affected tools & models
Never need to catch up again
The weekly delta: only verdict changes and act-now items. No digest filler.