- Published on
Gemini 3.8 Live Extended Thinking: The AI That Reasons While It Speaks
Imagine an AI assistant that interprets complex legal documents while holding a phone conversation.
It reviews case law while the other person is still talking, then delivers a precise conclusion the moment an answer is needed. That''s exactly what Gemini 3.8 Live Extended Thinking does — it maintains a real-time voice stream while running deep reasoning in the background.
1. How It Differs from the Standard Live Model
On September 15, 2026, Google released both Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking as generally available models.
| Feature | Gemini 3.8 Live | Gemini 3.8 Live Extended Thinking |
|---|---|---|
| Reasoning depth | Standard response | Deep (thinking_level: low / medium / high) |
| Async function calls | Not supported | Supported (tools run in background) |
| Best use cases | Fast replies, real-time translation | Complex problem-solving, multi-step logic |
| Speech-to-Speech rank | — | #1 on Artificial Analysis leaderboard |
| Input token limit | 131,072 | 131,072 |
The biggest differentiator is asynchronous function calls. The standard Live model pauses audio while a tool (search, calculation, etc.) executes. Extended Thinking keeps the voice stream alive while the tool runs in the background.
2. Setting thinking_level: How Deep Should It Think?
Control reasoning depth via an API parameter:
config = {
"thinking_config": {
"thinking_level": "high" # low / medium / high
}
}
low: Quick fact checks, simple calculationsmedium: Analysis that synthesizes multiple sourceshigh: Complex multi-step reasoning for legal, medical, or math problems
Note: The minimal option is not supported for the Extended Thinking model.
3. EdTech Use Cases
AI oral exam coach: The student speaks their answer; the AI analyzes the logical structure in real time and delivers specific feedback — not just "right or wrong," but "the causal link in this argument is weak."
Multilingual classroom assistant: Basic Gemini 3.8 Live already supports real-time interpretation in 97 languages. Adding Extended Thinking lets it reason about academic terminology context for more accurate translations.
Math tutoring in real time: The student explains their solution steps out loud; the AI listens step by step and pinpoints exactly where the reasoning goes wrong — the closest thing yet to a personal tutor.
4. Pricing and Access
- Pricing: 0.018/min audio output
- Free quota: Free tokens provided for both models at GA launch
- Model ID:
gemini-3.8-live-extended-thinking - Connection: WebSocket-based (not REST). Input: 16-bit PCM at 16 kHz; Output: 24 kHz
- Official integration partners: LiveKit, Pipecat, LangChain, Vercel, Agora
5. Practical Tips
1) Match thinking_level to your use case: low is enough for everyday conversational apps — it''s cheaper and faster. Reserve high for education, legal, or medical applications requiring deep reasoning.
2) Always use async tools: Any voice app that needs DB lookups or web search must use the Extended Thinking model. The standard model interrupts audio during tool execution.
3) Leverage integration partners: Pipecat and LiveKit let you build a voice agent in just a few lines of code without writing WebSocket connection code from scratch.
Source: