Google Launches Gemini 3.8 Live With Extended Reasoning for Voice Agents
Google launches Gemini 3.8 Live models with visual grounding, 97-language switching, background tools and extended reasoning for voice agents at scale.
Summary
Google launched Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026. Live targets fast, cost-efficient deployment, while Extended Thinking handles complex, multi-step tasks while speaking. Extended Thinking ranked first on Artificial Analysis’ Speech to Speech Quality Index with 82.6, scored 68.6% on τ-Voice, 35.1% on Sierra’s τ-Voice-banking benchmark and 97.7% on Big Bench Audio. Live placed second in Speech Agent Arena. Both advanced ServiceNow’s EVA-Bench Pareto Frontier for accuracy and conversational quality in testing through the Live API on Gemini Enterprise Agent Platform.
Both models process visual input in near real-time, switch automatically among 97 languages, and execute tools and API calls without stopping conversation. Extended Thinking acknowledges requests and narrates progress while completing background work, including React generation, bookings, onboarding, troubleshooting, chess, business plans and marketing materials. Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents support deployments, while Salesforce, Genspark and Lumeris are partners. Both are available through Gemini API and Google AI Studio. Live is reaching Search Live, with Gemini Enterprise in private preview and Customer Experience access coming soon. Extended Thinking is reaching Gemini Live, Docs for Google AI Pro and Ultra subscribers, and Gmail and Keep for all Google AI subscribers; Gemini Enterprise is in private preview, with Customer Experience and Google Workspace business access coming soon. All generated audio carries imperceptible SynthID watermarking.
Positives
- Gemini 3.8 Live Extended Thinking led Artificial Analysis’ Speech to Speech Quality Index with a score of 82.6.
- Extended Thinking reached 68.6% on τ-Voice, 35.1% on Sierra’s banking benchmark and 97.7% on Big Bench Audio.
- Gemini 3.8 Live automatically switches among 97 languages while maintaining an active conversation.
- Both models execute tools and API calls in the background without interrupting spoken interaction.
- SynthID imperceptibly watermarks all audio generated by Google’s AI products to support detection.
Risks & concerns
- Gemini Enterprise access begins in private preview rather than broad general availability.
- Customer Experience and Google Workspace business support is only described as coming soon, without release dates.
- Docs access requires Google AI Pro or Ultra, while Gmail and Keep require a Google AI subscription.
- Google calls pricing competitive and cost-efficient but provides no exact prices or comparative cost figures.
- ServiceNow EVA-Bench testing used the Live API on Gemini Enterprise Agent Platform, limiting the disclosed test context.