Google AI Updates: September 25, 2026

1. Gemini 3.8 Live Puts a Lip-Synced Avatar in Front of the Model

Google DeepMind released Gemini 3.8 Live with Live Avatar, a visual agent that handles video and audio input simultaneously while rendering a face with lip-syncing, expressions, and turn-taking. Tool calls run asynchronously so the avatar can keep talking while data is fetched in the background, and lip-sync adapts across 97 languages. Audio and video outputs carry SynthID watermarks. It ships in Gemini Enterprise, with custom avatars built from reference images gated behind an enterprise allowlist. The announcement gives no latency figures, benchmarks, or pricing. Source

2. The First Suncatcher Satellite Launches October 1 With TPUs On Board

Google is flying a refrigerator-sized satellite called MVP on a SpaceX Falcon 9 from Vandenberg on October 1, the first hardware test of Project Suncatcher, its plan to run solar-powered AI compute in orbit. The satellite was built and tested at Planet Labs in San Francisco and carries TPUs, and Google VP James Manyika says it should handle basic AI queries from orbit. The economics are not close yet: matching one terrestrial 1-gigawatt data center would take roughly 10,000 satellites, launch costs need to fall to about $200 per kilogram, and there is no atmosphere to cool the chips or shield them from cosmic radiation. Source

3. Call for Me Has Gemini Dial Businesses and Sit on Hold

Google started testing Call for Me, which has Gemini phone a business from the user’s own number to check stock, book a table, reschedule an appointment, or place a hold. Gemini introduces itself, navigates phone trees, waits on hold, and can share approved personal details, while the user follows a live transcript and can take the call over at any point. The test is limited to Pixel 11 owners in the US on a paid Gemini subscription running the beta Phone app, and Google says it is deliberately starting small because real conversations are nuanced. Source