Google AI Updates: September 24, 2026

1. Gemini 3.8 TTS Builds a Voice From a Text Description

Google. Gemini 3.8 Flash TTS and Flash-Lite TTS support more than 100 languages, accept stage directions on individual lines, and generate two-speaker dialogue from a single script. Flash TTS can synthesize a new voice from a text description of how it should sound, and a cloning feature builds a profile from a 30-second sample. Flash-Lite is positioned for high-volume generation with tone and pacing controls. Source

2. Private AI Compute Added Server-Side Memory That Google Says It Cannot Read

Google. DeepMind extended Private AI Compute with private server-side memory, letting personal AI retain context across sessions inside the same confidential compute boundary rather than pushing everything back to the device. The design point is persistence without Google gaining access to the stored context. Source

3. Antigravity SDK Can Run Agents Fully Offline on Gemma 4 26B

Google. The Antigravity SDK now supports local agentic execution through LiteRT with Gemma 4 26B A4B, plus drop-in support for any OpenAI-compatible server such as Ollama, LM Studio, or vLLM via LocalOpenAIAgentConfig. Google demonstrates an Architect-Builder split where Gemini 3.8 Flash plans in the cloud without uploading source and local Gemma instances do the work, reporting 97.2 percent of tokens in a security audit run staying on-device. Machines need more than 24GB of VRAM or unified memory. Source

4. XProf Opened Up Custom Pallas Kernels to Cycle-Level Profiling

Google. Google added a Kernel Profiling suite to XProf, its open-source TPU profiler, exposing cycle-level detail inside custom Pallas kernels. Those kernels previously showed up in trace captures as single opaque blocks, which made hand-written TPU kernels effectively unprofileable. Source

5. Made On YouTube Put Gemini in the Feed Algorithm and the Edit Timeline

Google. YouTube announced custom feeds that users describe in their own words and Gemini assembles, a storytelling assistant that analyzes scripts and rough cuts in Creator Studio, Gemini as a chat-based editor for Shorts, thumbnail performance tooling, and live English-to-Spanish stream translation. YouTube Music separately gained Ask Music for natural-language listening requests. Source

6. Gemini Omni Video Generation Landed in Google Vids for Free Accounts

Google. Google Vids now generates HD video through Gemini Omni, and Google is making it available to anyone with a Google account rather than gating it behind Workspace tiers. Source

7. Gemini Connected Apps Added Adobe, Airtable, Linear, and Peloton

Google. A new wave of Connected Apps extends Gemini’s reach into Adobe, Airtable, Linear, Peloton, and others. Linear in particular puts issue tracking inside the assistant’s tool surface, which matters for teams evaluating Gemini against Claude and ChatGPT connectors. Source