Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking launched on September 15 as Google’s new real-time audio models for conversational applications. The important change is architectural: the Extended Thinking model can keep a voice interaction moving while reasoning and non-blocking tool calls continue in the background.
- Google positions Gemini 3.8 Live for fast, high-volume conversations and Extended Thinking for multi-step voice work.
- Extended Thinking changes the session protocol: developers must watch an interaction-status signal instead of treating the end of one spoken turn as idle.
- For Indian support and commerce teams, the opportunity is less silence during tool calls, but the control burden rises when an agent can act across several systems.
Everyone else is reporting two new voice models; we are explaining why Gemini 3.8 Live forces developers to redesign the state machine behind a voice agent, not merely swap a model name.
What Gemini 3.8 Live launched
Google’s Gemini API documentation describes Gemini 3.8 Live Extended Thinking as an audio-to-audio model for complex, multi-step conversations. It accepts text, image, audio and video inputs, produces text and audio, supports the Live API and search grounding, and exposes a 131,072-token input limit with a 65,536-token output limit.
The standard Gemini 3.8 Live model is the lower-latency option. Extended Thinking adds configurable background reasoning and asynchronous function calling. Google’s launch claims include an 82.6 score on Artificial Analysis’ Speech-to-Speech Quality Index, 68.6% on the τ-Voice agent benchmark and 97.7% on Big Bench Audio. Those remain vendor-reported evaluation results, not guarantees for a specific deployment.
| Model | Best fit | Developer consequence |
|---|---|---|
| Gemini 3.8 Live | Direct, latency-sensitive conversations | Conventional turn completion remains the main boundary. |
| Gemini 3.8 Live Extended Thinking | Planning and multi-tool voice workflows | Apps must track background work until the session reports idle. |
Gemini 3.8 Live changes the control loop
Google’s Live API guide says Extended Thinking may speak an interim update, finish that utterance and then continue reasoning or calling tools. A client that treats turnComplete as the end of all work could enable new input too early, display a false idle state or lose a later tool request.
That distinction is the launch’s practical value. A travel or support agent can acknowledge a request while a reservation system, knowledge base or CRM responds. It is also the risk: every additional tool creates another permission boundary, timeout and audit event. The interface may feel more natural even as the backend becomes more complex.
What Indian enterprises should test first
India’s large contact-centre and IT-services workforce gives Gemini 3.8 Live an obvious proving ground. Teams should begin with bounded workflows such as order status, appointment changes or internal help-desk triage, then measure completion rate, interruption handling, tool errors and escalation quality in local accents and noisy environments.
The safer pattern mirrors the control work behind Zendesk’s specialised AI agents and Salesforce Koa’s CRM tool training: give each agent the minimum tool permissions, keep irreversible actions behind confirmation, and log the full chain from user request to tool result.
Independent coverage from SiliconANGLE reported the September 15 launch and highlighted the model’s real-time reasoning design. Vercel also listed both models in its AI Gateway on the release date, providing an observable downstream integration. Together with Google’s live documentation and model card, that supports the narrow claim that the models are launched and available—not that Google’s benchmark lead will transfer unchanged to every production workload.
Frequently asked questions
What is Gemini 3.8 Live?
Gemini 3.8 Live is Google’s low-latency real-time audio model for conversational applications. It is designed for fast dialogue, visual grounding and tool-enabled voice experiences.
How is Extended Thinking different?
Gemini 3.8 Live Extended Thinking can reason and make asynchronous tool calls in the background while continuing the voice interaction. Developers must track the interaction status until it reports idle.
Is Gemini 3.8 Live available to developers?
Yes. Google documents both models for the Gemini API and AI Studio, while its model card also lists distribution through Vertex AI and selected Google products.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
