Google Launches Gemini 3.8 Live and Extended Thinking Voice Models
Google launched Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, introducing near real-time parallel reasoning and background tool execution without halting conversational audio. Developers can build low-latency voice agents via the Gemini API, Google AI Studio, and streaming frameworks like LiveKit and Pipecat.
Impact: High
Why it matters
You can now wire asynchronous tool calls directly into live voice conversations without stalling the speech stream or waiting on user turn completion.
TL;DR
- 01Gemini 3.8 Live handles tool execution in the background without pausing conversational speech.
- 02Extended Thinking scores 82.6 on Artificial Analysis' Speech to Speech Index and 68.6% on τ-Voice.
- 03Both models are available immediately via Gemini API, Google AI Studio, and streaming kits like LiveKit and Pipecat.
Key facts
- Speech to Speech Quality Index
- 82.6 (#1 overall, self-reported)
- τ-Voice Agent Completion
- 68.6% (self-reported)
- Sierra τ-Voice Banking Benchmark
- 35.1% (self-reported)
- Big Bench Audio Score
- 97.7% (self-reported)
- Supported Languages
- 97 languages
Parallel Reasoning and Background Execution
Google has officially rolled out Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. The core architectural enhancement is the ability to process reasoning steps and function calls simultaneously with speech generation. Instead of waiting for a tool output to complete before synthesizing audio, Extended Thinking acknowledges user requests immediately using early verbal cues and narrates asynchronous background operations in real time.
Benchmark Results and Multilingual Capabilities
According to Google's reported metrics, Gemini 3.8 Live Extended Thinking achieved the #1 spot on Artificial Analysis' Speech to Speech Quality Index with a score of 82.6. In agentic task completion, it scored 68.6% on τ-Voice and 35.1% on Sierra's τ-Voice-banking benchmark, alongside 97.7% on Big Bench Audio. The standard 3.8 Live model supports visual grounding and automatic language switching across 97 languages mid-conversation.
Developer Tooling and Availability
Developers can access both models through the Gemini API and Google AI Studio. Real-time media streaming integrations are natively supported via partners including Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, and Vision Agents. All generated speech streams include imperceptible SynthID audio watermarking.
✓ When to use
- Building interactive customer support or operational voice agents requiring parallel backend lookups.
- Implementing hands-free multimodal troubleshooting where users stream camera feeds and speak simultaneously.
✕ When NOT to use
- Do not use for text-only pipeline architectures where standard asynchronous REST or batch APIs are significantly cheaper.
- Avoid for ultra-strict zero-hallucination compliance without human verification in high-risk domains.
What to do today
- Test low-latency voice agent workflows inside Google AI Studio or through the Gemini API.
- Review LiveKit or Pipecat integration docs for streaming real-time audio sessions to Gemini 3.8 Live.
What the community says
“I had a funny moment with ChatGPT. I was talking to it in English and then asked it to speak Swedish with me. It did, but with an English accent. But if you just start a session in Swedish, it speaks Swedish perfectly with local accent”
“It is by far the least accurate of any model I have used too. For a company that was started to organize the world's information, it has by far the most misinformation I've encountered.”
Sources