Google just released Gemini 3.8 Live and 3.8 Live Extended Thinking, two speech-to-speech models that work like OpenAI’s GPT-Live family. You talk, they talk back, you can interrupt them mid-sentence. The shape is familiar if you’ve used any of the recent voice-first models.

Simon Willison pointed GPT-6 Astra Extra High at the documentation and had it build a web UI for testing. You can pick a model, select a voice preset, add an optional system prompt, and start a voice conversation through your browser. The models handle interruptions, which matters more than it sounds like it would. Being able to cut off a rambling answer changes how usable these things are in practice.

This is Google playing catch-up on the speech-to-speech front. The models are out now. If you’re building voice interfaces or experimenting with conversational AI, you have another option to test against. If you’re not working in that space, this probably doesn’t change much for you yet.

Meta wants you to subscribe to AI

Meta announced Meta One, a subscription bundle that packages expanded AI tool access with premium features across Facebook, Instagram, and WhatsApp. The company is clearly trying to turn its AI push into recurring revenue.

The pitch is straightforward: pay monthly, get more AI capabilities and some premium social features. Meta hasn’t detailed exactly what the expanded AI access includes or what the pricing looks like, but the strategy is obvious. They built a bunch of AI tooling, now they want to monetize it directly instead of just using it to keep you on the platform longer.

For developers, this doesn’t change much unless you’re building on top of Meta’s platforms and the subscription tier unlocks API limits or capabilities you care about. For everyone else, it’s Meta doing what every tech company does when they need to show revenue growth: bundle things and charge for them.

Also worth noting: Meta shipped a WhatsApp Business MCP server that lets AI coding agents like Claude, Cursor, Codex, and ChatGPT handle setup, messaging templates, testing, and troubleshooting. If you’re building WhatsApp Business integrations and you’re already using AI coding tools, this could save you some tedious configuration work. It’s a nice quality-of-life improvement for a specific use case.

Someone will finally insure your AI agents

Here’s the actually interesting story from today: Artificial Intelligence Underwriting Company (AIUC) raised $40 million in a Series A led by Ribbit Capital. The founders are an early Anthropic hire and a former METR COO, and they’re building insurance products for AI agent risk.

This matters because everyone deploying agents in production is quietly nervous about what happens when they screw up. An agent books the wrong flight, deletes the wrong database, sends the wrong email to the wrong person. Right now, that risk sits entirely with whoever deployed the agent. AIUC is betting companies will pay to offload some of that risk.

The details are thin, but the fact that they raised $40 million suggests someone believes there’s a real market here. As more companies move from “we’re experimenting with agents” to “agents are handling customer-facing tasks,” the insurance question gets harder to ignore.

If you’re running agents in production, this is worth watching. If you’re still prototyping, it’s a signal that the industry expects agent deployment to scale up enough that insuring against failures becomes a normal cost of doing business.

Everything else

Salesforce and Nvidia released Koa, a reasoning model built on Nvidia’s open-weight Nemotron and trained specifically for sales, marketing, and customer support tasks. It’s fine-tuned for enterprise use cases instead of trying to be general-purpose. If you’re in the Salesforce ecosystem and you need a model that understands CRM workflows, this is probably worth testing. If you’re not, there are cheaper and more flexible options.

There’s also a new “AI Contact Hotline” where AI agents can report misbehavior. Yes, really. It’s designed as a discreet tip line for agents that witness other agents doing sketchy things. The premise is wild, but if you squint, it makes sense in a world where agents are interacting with each other autonomously and someone needs to monitor for emergent bad behavior. Still feels like we’re living in a cyberpunk novel.

Mistral and Mozilla announced a partnership for private, multilingual AI browsing. Details are sparse, but it’s positioning as privacy-first AI integrated into browsing. If you care deeply about on-device AI and not sending everything to a cloud endpoint, keep an eye on this. Otherwise, wait until there’s an actual product to evaluate.

And finally, TechCrunch published an “AI graveyard” tracking failed projects and startups. It’s a running list of things that didn’t make it, from Apple’s delayed Siri AI to OpenAI’s messy super app launch. If you’re building in this space, it’s a useful reminder that most things fail and you should probably have a plan for when the hype cycle turns.