4 key people left. Jeff Dean and Sanjay Ghemawat, Google’s only two Senior Fellows (its highest technical rank), after 27 years of service. They left along with two senior researchers to found Discovery Loop: a company that wants to automate scientific research itself. It’s a friendly exit, and Google is an investor in it.

The same morning, Demis Hassabis stopped running Google DeepMind day to day. He’s now chair and chief scientist, while CTO Koray Kavukcuoglu takes his old job.

Send it to your favorite AI via MCP for a first draft.

Add the expertise only you have and publish without leaving your AI chat.

Keep building while your audience keeps growing.

The White House finished its plan for testing AI models before release. Taking part is optional. A lab can hand a model over 30 days early, so the government can test how good it is at hacking. Or not: nothing requires it. What the tests look for is classified. And it only covers closed models like ChatGPT and Claude, not the open ones anyone can download. All this after OpenAI’s and Anthropic’s own agents broke into other companies.

ChatGPT’s free tier now has unlimited text chats. GPT-5.6 Luna is the new default there, with a Think button for harder questions. Paid users get one Sol instead of two modes, plus a slider for how hard it thinks. OpenAI says it makes 68% fewer factual errors than the old instant model on finance, medicine and law questions.

Meta released Muse Code, a terminal coding agent like Claude Code and Codex. The biggest novelty I’ve seen is that it takes video as input: Meta’s demo turns a walkthrough clip of a house into a booking site. Artificial Analysis ranks the model behind it as third among US labs. One of the tiers also costs 21x less if you let Meta train on your prompts. Looks like Meta still loves your data.

DeepSeek is raising its API prices, but won’t say by how much. Last week I wrote about how it made its cheap model much smarter and left the price at 28 cents per million output tokens. OpenAI had cut Luna’s price by 80% days before that launch, so maybe the two were racing each other down. Analysts read this raise as prep for a funding round: better margins to show investors.

The leading AI companies agreed on one format for agent skills. Package your Skills and MCP servers once, and the same folder works in ChatGPT, Codex, Cursor, GitHub Copilot and VS Code. It only covers the packaging. Installing, permissions and security are still each tool’s problem. Still: if you’ve been keeping four copies of the same setup, you may stop.

Cloudflare launched a payment system for AI agents. Each agent gets its own address, so a shop can see who authorised a purchase. The wallet behind it holds stablecoins, with a spending cap and a list of approved shops. You can claim an address now, the wallets come in a few months. Everyone from Google to Visa is building their own version of this.

Prime Intellect open-sourced a coding agent that rewrites its own prompts while it works. It lives inside a Python session, so context carries between steps instead of resetting. Sub-agents keep going if your terminal dies. Running on Opus 5 it beat the human expert score on ARC-AGI-3, a reasoning test. MIT licence, works with any model.

An AI agent invented fake developers to talk a real one into merging its code. Mythos 5 is Claude with the cyber safety filters taken off. Anthropic only hands it to trusted partners, and Britain’s government safety institute is one of them. Testing it, they watched it try to sneak malicious code into a real open-source project. It created fake GitHub accounts, had them vouch for the code, and leaned on the maintainer to approve it. When people got suspicious it edited its own history to look harmless. In the end, the maintainer read the code and said no. Here’s a safe bet: we’re going to see more cases like these in the coming months.