Google launches Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, new near-real-time voice dialogue models for Gemini Live and the Gemini API

Decision brief

What changed

Google introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, new near-real-time speech-to-speech dialogue models for production voice agents. Google made them available through the Gemini API and related Google platforms, adding capabilities such as streamed conversation, multilingual switching, interruption handling, and background tool calls.

Why it matters

This gives enterprises a new Google-native option for lower-latency voice interfaces that can support more natural, continuous conversations and tool-enabled workflows. For leaders deciding on customer service, productivity, or assistant experiences, the launch expands the practical set of deployable voice-agent models and introduces usage-based audio pricing that may affect build-vs-buy and platform selection decisions. The broader rollout across Google platforms also matters for organizations already committed to Google’s ecosystem.

Affected roles

CEO CFO COO CTO CMO

Evidence

The core announcement comes from Google DeepMind, which details the two models, their features, platform availability, and the Artificial Analysis benchmark result. Independent coverage from MarkTechPost, Simon Willison, and SiliconANGLE is broadly consistent on the launch, near-real-time voice focus, API availability, and cited pricing/technical access details.

What remains uncertain

The coverage does not verify real-world performance, reliability, or total operating cost in production deployments beyond Google’s claims and a third-party benchmark score. It also remains unclear how the models compare on enterprise-specific needs such as accuracy across accents, compliance requirements, and integration complexity outside Google’s own platforms.

Monitor next

Watch for broader enterprise availability details, customer deployment references, and independently reported latency, quality, and cost results versus competing real-time voice models.

Analytical support, not advice — assumptions and open questions stated above.

Source coverage