Google launched Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026. The new live dialogue models process visual inputs in near real time, support 97 languages with automatic switching, and execute background tool calls while maintaining natural conversational flow.
Voice assistants have long suffered from an awkward conversational trap: whenever a user asks a complex question, the system falls completely silent while it fetches data or runs a tool. Google is attempting to fix that exact latency gap with a pair of native speech-to-speech models designed to reason and speak simultaneously. Introduced by principal engineer Tom Ouyang and technical staff member Malini Jaganathan on behalf of the Gemini Audio Team, the release spans the Gemini API, Google AI Studio, Gemini Enterprise, Search Live, Gemini Live, and Google Workspace.
Dual-Model Architecture for Scale and Complex Tasks
The rollout divides workloads between two distinct models tailored for different operational demands. Google describes the pair as its most advanced live dialogue systems to date, built to handle fluid conversation alongside parallel reasoning for demanding workflows.
Gemini 3.8 Live is engineered for scale and cost efficiency, pairing conversational intelligence with visual grounding and rapid response times. It secured second place in Artificial Analysis’ Speech Agent Arena, backed by strong user preference scores. Meanwhile, Gemini 3.8 Live Extended Thinking targets high-complexity tasks that demand multi-step reasoning, background planning, and deeper analysis.
| Feature | Gemini 3.8 Live | Gemini 3.8 Live Extended Thinking |
|---|---|---|
| Primary Use | Low-latency, real-time voice interactions | Complex, multistep tasks requiring deep reasoning |
| Tool Execution | Async function calling by default | Async, nonblocking function calls only |
| Key Benchmarks | 2nd in Speech Agent Arena | #1 on Speech-to-Speech Quality Index (82.6); 97.7% on Big Bench Audio |
Background Reasoning and Asynchronous Tool Execution
The technical architecture changes how voice agents handle multi-step actions. Rather than blocking the conversation while executing an API call, both models rely on asynchronous function calling. The system acknowledges user requests instantly, keeps the audio stream open, and processes tools in the background.
When handling complex problems, Extended Thinking utilizes early verbal cues to signal it is working on a request.
For developers building production-grade applications, Google’s Live API documentation stresses that developers must track interaction states rather than relying solely on traditional completion signals. A turn-complete signal can arrive after an intermediate spoken update while background reasoning continues, meaning downstream systems must monitor state changes before treating a task as finished.
Ecosystem Integration and Developer Pricing
The models are accessible to developers through the Gemini API and Google AI Studio. Infrastructure partners including LiveKit, Vercel, Agora, Fishjam, Pipecat, and Vision Agents provide real-time media streaming support for voice-driven interfaces. Enterprise adopters such as Salesforce, Genspark, and Lumeris are also integrating the technology.
Google has established a standardized pricing structure for developers across both models. Audio input is priced at $3 per million tokens, translating to approximately $0.005 per minute, while audio output is set at $12 per million tokens, or about $0.018 per minute. For consumer rollouts, Gemini 3.8 Live is deploying via Search Live, Extended Thinking is rolling out across Gemini Live, and Google AI subscribers gain workspace access in Docs, Gmail, and Keep.
Multilingual Capabilities and Production Guardrails
Beyond speed and reasoning, the models process near-real-time visual inputs and automatically detect and switch between 97 supported languages mid-conversation. Alphanumeric precision has also been tuned to accurately parse technical data, confirmation codes, and claim numbers—areas where traditional voice systems frequently fail.
