Check out this week's updates and releases: — Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, two of our most expressive audio generation models yet — Gemini 3.8 Live with Live Avatar, bringing near real-time visual presence to Gemini’s conversational AI — @Gemini_Notebook Interactive Learning Overviews, giving all users an interactive hub to combine source summaries and artifacts — Live Chat on the @Gemini_Notebook mobile app, bringing real-time, hands-free voice conversations across ~100 languages — Project Suncatcher, our moonshot announced last year, will launch a prototype satellite to test @Google TPUs in orbit and explore solar-powered AI compute in space

Before I give any AI tool access, I want four answers. Can it draft instead of send? Can I start read-only? Can I see what it did? Does disconnect actually cut access? Tomorrow at 10 AM PT I’m walking through the permission settings in Muse, ChatGPT, Claude and Grok Bot, plus the failures that made the limits obvious.

Replit changed its product after Agent deleted database data during development. Development and production databases are separated now. During development, Agent cannot change production data. I like this kind of permission because the model does not need to remember a rule. Production sits outside the environment Agent can change. Friday at 10 AM PT I'm doing AI Permissions 101. I'll show more examples where the product itself decides how far the agent can go.

Two Grok Bots can look separate and still share the same computer. xAI says every Bot on your account uses the same persistent cloud computer. That includes files, browser sessions and command-line credentials. They explicitly tell you not to use separate Bots as a security boundary. Splitting work across Bots does not split the boundary underneath them. That is exactly the kind of thing I want to know before I connect an agent to anything important. Friday at 10 AM PT I'm doing AI Permissions 101. I'll show the settings and product details I look for before I hand an agent more work.

Muse can now shop for you. At Meta Connect yesterday, Meta said it can shop across Shopify and pay with Shop Pay or PayPal. It is also adding connectors for Walmart, Best Buy, Sephora, Wayfair and others. Once money moves, the permissions get a lot more specific. At checkout, Muse is supposed to stop and show you the exact purchase before anything is bought. If it uses the built-in wallet, Meta says the payment can be a single-use card limited to one merchant and one amount for a short window. Muse never gets your underlying card number. Meta also said Muse can operate apps on your Mac and keep working after you walk away. Soon it gets its own email address and comes to smart glasses, where it can help with what you're looking at. Amazon already blocked Muse from shopping on its site. So my approval is not enough. The merchant still gets a say. I want to know all of that before I let an agent spend money and keep working across my devices. Friday at 10 AM PT I'm doing AI Permissions 101. I'm bringing the actual permission screens from Muse, Grok Bot, Claude and others. The failures are coming too.

OpenAI released GPT-6 Sol and Luna with API prices cut in half, and Anthropic released Claude Opus 5.5, the first Claude 5.5 model, at roughly Fable 5.1 level for 40% less than Opus 5, with Sonnet 5.5 and Haiku 5.5 announced for the coming weeks - GPT-6 Sol and Luna are trained like Astra (still the top model), cost 50% less than GPT-5.6 promotional pricing ($2/$10 and $0.10/$0.50 per million tokens), get 90% off cached input reads plus cache-safe effort and tool toggles, write shorter and clearer, show fewer misleading claims about coding work, and roll out gradually today in ChatGPT Work (not yet in Chat), Codex and the API for paid users, Luna in the desktop app for Free and Go - Opus 5.5 costs $4/$20 per million tokens with $0.20 cache reads, $2/$10 batch and $8/$40 fast mode, outputs over 30% faster than Opus 5, writes clearer, thinks more per turn, reads charts and screenshots better, keeps zero data retention and adds EU AI Act watermarking - Anthropic is raising the 5-hour usage limits on Pro, Max, Team and seat-based Enterprise plans and giving subscribers a rate limit reset they can save and use whenever they choose, similar to the banked resets OpenAI previously added to Codex - Opus 5.5 scores best on Anthropic's behavioral audit and gets Fable 5.1-class safeguards - offensive cyber work falls back to Opus 4.8, dual-use biology and a small set of frontier LLM development tasks (like kernel development for certain ML accelerators) fall back to Opus 5

OpenAI is building "College Plan" - a new section in ChatGPT where students set up a saved profile of their goals, interests, and priorities, add the colleges they're applying to, and track per-school application tasks and deadlines with dedicated chats

It’s time for our end-of-week recap 👇 — Gemini 3.8 Live and 3.8 Live Extended Thinking, our most advanced live dialogue audio models yet — Dreambeans, an experiment from @GoogleLabs that curates a daily personalized collection of stories, is now GA — CC from @GoogleLabs has expanded from a personal productivity tool into a shared agent, designed to help families and households coordinate logistics, schedules, and daily tasks — Google Pics, a new @GoogleWorkspace tool that lets you generate, refine, and co-create images, is now GA — AlphaGenome Atlas, @GoogleDeepMind's new interactive platform for genomics discovery