Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking

What it claims

Google DeepMind released two speech-to-speech models on the Gemini Live API.

• Gemini 3.8 Live is the cost-efficient, high-volume tier.
• Gemini 3.8 Live Extended Thinking is the reasoning tier for multi-step agent workflows.

Beyond the headline scores, the page reports 35.1% on Sierra's τ-Voice-banking for Extended Thinking. It also claims both models "push the Pareto frontier" on ServiceNow's EVA-Bench, balancing accuracy against conversational quality on complex workflows. That EVA-Bench result was run on the Live API on Gemini Enterprise Agent Platform.

How it works

The page describes behaviour, not architecture.

3.8 Live:

• Takes visual input in near real time.
• Detects and switches language automatically mid-conversation.
• Runs tool and API calls in the background. The model acknowledges a request and keeps talking while the call completes, so your agent does not have to hold a silent turn during tool latency.

3.8 Live Extended Thinking reasons and speaks at the same time:

• It gives early verbal acknowledgements such as "Let me check that…" to cover reasoning time.
• It narrates progress aloud while multi-step background tasks run.

Engineers should expect these filler phrases and narration in transcripts and in the user experience.

How to use it

Developer access is through the Gemini API and Google AI Studio, rolling out from the announcement date for both models.

Enterprise access:

• Both models are in private preview in Gemini Enterprise.
• Gemini Enterprise for Customer Experience is listed as "coming soon".
• Workspace business customers are listed as "coming soon" for Extended Thinking only.

Consumer surfaces:

• 3.8 Live is in Search Live.
• Extended Thinking is in Gemini Live.
• In Workspace, Extended Thinking reaches Docs for Google AI Pro and Ultra subscribers, and Gmail and Keep for all Google AI subscribers.

The page names third-party platforms that already wrap the Live API and handle real-time media streaming: Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, Vision Agents. If you already use one of these, it is the likely integration path. Named early partners are Salesforce, Genspark and Lumeris.

All generated audio carries a SynthID watermark. You cannot opt out of it according to the page ("All audio generated by our AI products").

Limits

• No prices or latency figures. Cost is described only as "highly cost-effective" or "highly competitive", with no per-token or per-minute numbers.
• Benchmark scores have no baselines. No competitor scores are shown.
• The Speech Agent Arena entry does not say who ranked first.
• The EVA-Bench claim has no numbers in the text.
• Model identifiers are not given. The page has no API model strings, context limits or rate limits.
• Safety detail is elsewhere. Safety and responsibility details are deferred to a separate model card, which is not included here.
• Enterprise availability is limited. Enterprise use is private preview only, so check access before planning production deployment there.