Google DeepMind adds real-time avatar video ("Live Avatar") to Gemini 3.8 Live for Gemini Enterprise

Google DeepMind added "Live Avatar" to its Gemini 3.8 Live model, generating a near real-time animated avatar (lip-sync, expressions, turn-taking) alongside voice output, with lip-sync switching across 97 languages and background tool-calling that runs without interrupting conversation. It is live today inside Gemini Enterprise only, with a preset avatar library and an allowlisted option for custom avatars built from a reference image. Outputs carry a SynthID watermark. No pricing, latency figures, or wider availability date have been given.

This affects enterprises building customer service bots or interactive walkthroughs on Google's stack, giving them a visual layer to pair with voice without building their own avatar pipeline. For most professionals the immediate impact is limited: it's gated to Gemini Enterprise customers, custom avatars need allowlisting, and there's no consumer or general API rollout yet, so this is a feature for a narrow, already-committed buyer base rather than a broad market shift. The multilingual lip-sync and async tool-calling are genuinely useful engineering details if true, but there's nothing here yet that changes competitive dynamics.

"Near real-time" is never quantified, so there's no way to check the actual latency claim, and phrases like "richer, more accessible experiences" are marketing language with no data behind them. This is also framed as a breakthrough when real-time avatar agents already exist commercially from HeyGen, Tavus and others, so the novelty is mainly "Google now has one too," not a new capability category.

1. 01Whether Live Avatar reaches the general Gemini API or consumer products, and when, versus staying locked to Gemini Enterprise
2. 02Any published latency, cost-per-minute, or benchmark comparison against HeyGen/Tavus avatar agents once third parties test it
3. 03Enterprise adoption signals: named customers, case studies, or usage numbers in the next one to two quarters rather than just the launch post