Google unveils Gemini 3.8 Live and Live Extended Thinking speech‑to‑speech models

Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, adding real‑time speech‑to‑speech capabilities that developers can reach through a direct WebSocket endpoint and the Web Audio API AudioContext.

Key points

• Gemini 3.8 Live and Live Extended Thinking are speech‑to‑speech models.
• The models mirror the shape of OpenAI’s GPT‑Live family.
• Simon Willison built a no‑library web UI to demo the models.
• Interaction occurs over wss://generativelanguage.googleapis.com/ws/... endpoint.

What happened

Simon Willison pointed GPT‑6 Astra Extra High at Google’s documentation and generated a browser‑based UI that lets users select a model, choose a voice preset, and provide an optional system prompt. The interface also supports interrupting the model while it speaks, delivering a fluid conversational experience.

The UI connects straight to the GenerativeService BidiGenerateContent WebSocket, feeding audio captured via a Web Audio API AudioContext and streaming the model’s responses back through the same context. No external libraries are required, demonstrating that the core functionality can be built with native browser APIs alone.

The backstory

Speech‑to‑speech generation has accelerated since OpenAI introduced its GPT‑Live series, which allow real‑time voice interaction without intermediate text steps. Google’s Gemini line has historically focused on multimodal text and image generation, but the company has hinted at expanding into live audio processing in prior roadmap briefings.

Industry observers note that the shift toward low‑latency voice agents reflects broader demand for hands‑free interfaces in customer support, accessibility tools, and IoT devices. Google’s move aligns with that trend, positioning Gemini as a direct alternative to OpenAI’s live offerings.

Competitive implications

By matching the architecture of GPT‑Live, Gemini 3.8 Live narrows the functional gap between Google and OpenAI in the emerging voice‑agent market. Enterprises that already use Google Cloud services may adopt the new models to avoid multi‑cloud complexity, while developers seeking open‑source‑style flexibility gain a ready‑made WebSocket endpoint.

Technical trade‑offs

The reliance on a raw WebSocket and the Web Audio API offers maximum control but places latency management and audio quality squarely on the client side. Teams must handle buffering, echo cancellation, and network variability themselves, unlike higher‑level SDKs that abstract those concerns. The trade‑off is lower overhead and tighter integration with existing web stacks.

What builders and businesses should do

• Test Gemini 3.8 Live via the public WebSocket endpoint to benchmark latency against GPT‑Live.
• Use Simon Willison’s demo UI as a reference implementation for rapid prototyping.
• Evaluate audio handling requirements, especially echo suppression, before production deployment.
• Monitor Google’s pricing and quota policies for the live endpoint to forecast operating costs.
• Integrate voice‑interrupt capabilities early to preserve conversational fluidity.

The bottom line

Watch for Google’s pricing announcement and broader SDK support in the coming weeks; those signals will determine how quickly enterprises can adopt Gemini Live for real‑time voice applications.