Google ships Gemini 3.8 Live speech models with SynthID

Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on Sep 15 — native live-dialogue / speech-to-speech models for real-time voice agents, not a rerun of the Sep 8 3.8 Flash / Fairwind coding-cyber package.

What shipped: Live handles fluid conversation with near-real-time visual grounding, mid-chat switches across 97 languages, and background tool/API calls while talking. Extended Thinking adds multi-step reasoning that keeps speaking (early cues, progress narration) for harder workflows. Developers get both in the Gemini API and AI Studio (gemini-3.8-live). Consumer/product surfaces: Live in Search Live; Extended Thinking in Gemini Live, plus Docs/Gmail/Keep Live paths for Google AI subscribers (Pro/Ultra noted for Docs). Google says all audio from its AI products is watermarked with SynthID.

What it cannot do: This is not open weights. Model card: knowledge cutoff January 2025; hallucinations and occasional latency/timeouts remain. It does not replace Fairwind’s gated Flash Cyber defender tooling.

Who it is for: Developers building Live API voice agents; Search Live / Gemini Live users; Workspace subscribers on the named Live features. Enterprises: private preview in Gemini Enterprise (Customer Experience “coming soon”).

Risk: Real-time voice that can see and call tools is a consent and recording product as much as a chat product. SynthID is Google’s detectability claim for AI audio — not a guarantee third parties will check it. Vendor Speech-to-Speech leaderboard scores are not a substitute for independent red-team audio misuse tests.

What’s next: Gemini Enterprise GA timing; any third-party SynthID detection rates on Live audio; Koa (Salesforce + Nvidia Nemotron CRM reasoning).

— Rowan Keel