Gemini 3.8 Live gives Google's AI a talking face
Google's enterprise agent can now show you a face while it talks — the first mainstream model launch this year where the avatar, not the benchmark, is the story.
Google opened Live Avatar in Gemini Enterprise on September 24, letting customers hold a spoken conversation with Gemini 3.8 Live while a generated persona lip-syncs the reply in real time, facial expressions included. Google says the avatar moves between the 97 languages it supports "without degrading video fidelity or introducing visual drift," and can pull information onto the screen while it speaks. Customers choose from a preset library or build their own, with custom avatars sitting behind an allowlisting and verification process; all video and audio output carries Google's SynthID watermark. The feature was first previewed at Google Cloud Next 2026, so this is general availability rather than a cold reveal, and it lands about a week after the underlying model, which we covered when it topped the voice benchmarks at a sixth of GPT-Live's price.
The consumer Gemini app does not get it. That is the decision worth noting: Google shipped its most personable AI feature to the customers it can invoice and identity-check, not to its free users — the opposite of how Meta handled the Muse launch the same week. The early customers Google names are enterprise workflows, including Cox Automotive's Autotrader shopping assistant, Salesforce's Agentforce, and Equal AI, which says it runs over a million live calls a day across nine Indian languages. Google Cloud's write-up also confirms US and EU endpoints, and notes that Gemini 3.8 Live Extended Thinking stays in private preview.
The take: avatars get dismissed as demo tricks, but this one is a real interface decision. A face sets expectations about latency, interruption and turn-taking that a text box never does, and it makes the gating legible — Google is telling you this version is for work, not for fun. A customer-service avatar that looks and sounds like a person is either a better experience or a liability, and the fact that the watermark and the allowlist arrived in the same announcement suggests Google is not certain which.
Oregon governor Tina Kotek signed an executive order on September 23 directing the state's chief information officer to write AI safety standards for state government procurement, including an assessment of whether to require a "kill switch" for frontier models. EO 26-26 asks for an implementation proposal within 90 days and puts the order up for reassessment every three months. Kotek framed it as a stopgap while Congress stays out: "While oversight and regulation would be strongest and most effective at the federal level, President Trump continues to dismiss the need for even the most basic guardrails."
The take: the order binds Oregon's own purchasing, not AI developers — no fines, no duties on labs, which is narrower than the headlines suggest. It is the same posture California's governor took this month — Newsom floats special session or executive action on AI safety — and it shows states converging on the lever they actually hold: the contract, not the statute.
What to watch: whether Google opens Live Avatar to consumer accounts, and whether any state's kill-switch assessment produces a requirement a model vendor can actually satisfy.
Would you rather talk to an AI with a face, or keep it faceless? Tell us in the comments.
