Gemini3.8 Live
#105 in Developer Tools
« A real-time voice dialogue model that reasons during the conversation, processes visual inputs, and executes tool calls in the background »
Verified Tool
Gemini 3.8 Live: talk to Google's AI while it works behind the scenes
Gemini 3.8 Live is Google DeepMind's real-time voice dialogue model, paired with an Extended Thinking variant for complex, multi-step tasks. Both arrived on September 15, 2026 and can run tool and API calls in the background without breaking the conversation. 97 languages are detected and switched on the fly, mid-sentence if required. Access runs through the Gemini app, Search Live, Google Workspace and the Live API for developers.
Pros
- First place on the Speech to Speech Quality Index
- Tool calls executed while the conversation keeps going
- 97 languages detected and switched mid-call
- Camera feed analyzed in near real time
- Already available through the Live API and Google AI Studio
Cons
- No detailed API rate card at launch
- Hosted models only, no open weights
- Full access in the Gemini app tied to paid plans
Gemini 3.8 Live speaks, watches and calls tools without dropping the thread
Gemini 3.8 Live processes speech end to end, skipping the usual chain of speech recognition, language model and text-to-speech. Google DeepMind shipped two versions on the same day. Live targets scale and contained costs, while Extended Thinking absorbs complex jobs with multi-step reasoning carried out while it answers you.
In practice, you point the camera at an invoice, the agent reads it, queries an API in the background and finishes its sentence without a gap (that awkward silence of a voice assistant lost in thought, everyone knows it). Interruptions don't throw it off either.
- Continuous dialogue that survives interruptions
- 97 languages detected on the fly
- Tool and API calls handled in the background
- Camera feed and shared images read in near real time
Top of the voice benchmarks, with numbers to back it
Gemini 3.8 Live Extended Thinking holds the number one spot on Artificial Analysis Speech to Speech Quality Index with 82.6, ahead of GPT-Live-1 Astra and Grok Voice Think Fast 2.0. The lighter Live model took second place in the Speech Agent Arena, a ranking built on human preference votes.
On Sierra's banking tasks, 35.1% still leaves headroom. An honest sign that voice agents remain young on regulated workflows, even at the top of the table.
| Benchmark | Gemini 3.8 Live Extended Thinking | Closest rival |
|---|---|---|
| Speech to Speech Quality Index | 82.6 (1st place) | GPT-Live-1 Astra, 81.5 |
| τ-Voice (agent tasks) | 68.6% | GPT-Live-1 Astra, 67.9% |
| Sierra τ³-Banking | 35.1% | GPT-Live-1 Astra, 32.0% |
| Big Bench Audio (reasoning) | 97.7% | not disclosed |
Where to try Gemini 3.8 Live, from Search Live to the Live API
Both models can be called today through the Gemini Live API and Google AI Studio, with LiveKit, LangChain and Agora already plugged in to speed up deployments. Companies get a private preview on Gemini Enterprise. On the consumer side, Search Live serves them now, while the Gemini app ties them to Google AI Pro at $19.99 per month and Ultra from $99.99. Google published no separate rate card for these models at launch, while claiming a competitive position against other frontier models. Those figures will shift without warning, so budget from the official pricing page and nowhere else.
The race between autonomous voice agents is now decided on latency and reliable tool calling, and Google is fielding its strongest cards here.
Frequently asked questions
Is Gemini 3.8 Live free?
Partly. Search Live already runs the model at no extra charge, and Google AI Studio lets you test it for free. In the Gemini app, the new voice models sit behind the Google AI Pro and Ultra plans, and production access through the Live API bills per usage.
What is the difference between Gemini 3.8 Live and Extended Thinking?
Two distinct roles. Gemini 3.8 Live aims at scale and cost control, with fluid dialogue and visual grounding. Extended Thinking adds background multi-step reasoning for complex scenarios such as technical diagnostics or bookings spanning several APIs, without breaking the rhythm of the call.
Can you self-host Gemini 3.8 Live?
No. Both models are hosted by Google and reached through the Live API, Google AI Studio or the Gemini Enterprise preview. No open weights are distributed, unlike the Gemma family. Your audio therefore passes through Google's infrastructure, a point worth checking for regulated industries.
How many languages does Gemini 3.8 Live support?
Ninety-seven languages are covered, and the model detects which one you are speaking automatically, even when it changes mid-call. An agent can open in English and switch to Spanish or Hindi without any setup, a strong argument for multilingual customer support lines.
Verdict: For phone desks, multilingual support lines or voice tutoring, Gemini 3.8 Live fills the missing slot, and voice agent developers along with Gemini subscribers gain a conversation partner that never loses the thread.
★ Featured AI Tools ★
AI Alternatives for
Gemini3.8 Live
Freemium
Free
Free
Free
Free-Trial
Free
Free
