Gemini 3.8 Live Avatar goes GA: ship a real-time video agent
Google's speech-to-speech model now drives a lip-synced avatar that sees your screen — plus OpenAI's GPT-6 caching rework and Gemini voice cloning.

Copy markdown
A real-time video agent you can actually ship
Gemini 3.8 Live with Live Avatar is now generally available on Google Cloud — a native speech-to-speech model that renders a lip-synced video avatar, reads live camera and screen feeds, and calls your tools mid-conversation across 97 languages. Prebuilt avatars work today through the Live API and the Agent Development Kit; custom avatars stay enterprise-allowlisted, and every clip carries a SynthID watermark.
GPT-6 prompt caching gets a dashboard and breakpoints
OpenAI reworked GPT-6 caching: cached input still earns up to a 90% discount, but hit rates are higher by default and you now get a caching dashboard, a diagnostics tool that names miss reasons like 'tools_changed,' explicit cache breakpoints, and prewarming. You can shift reasoning effort or swap tools mid-conversation without busting the cache — real agents are reporting cache hits climbing from ~83% to 91% in under a week.
Gemini's text-to-speech can now clone a voice
Gemini 3.8 Flash TTS and Flash-Lite TTS are GA with voice design for custom personas, consent-verified voice replication, and 150+ voices in the Extended Voice Library — enough to stand up a branded narrator or a dubbing pipeline straight from the API.