| Good morning. It’s Friday, November 13th. | Just a quick note to my hardcore Anthropic users: Claude Connectors appear to be offline this morning - so if you’re wondering why your MCPs are asking you to reconnect, it’s on Anthropic’s side. | -Jeff AI Breakfast |
|
| | | In partnership with Descope | Your users’ AI agents are already trying to log in. Envoy lets them in safely. | | Personal agents like Muse and Instinct are already visiting your website. Blocking them blocks revenue, and letting them log in as the user causes security issues. | Envoy from Descope gives you a third path: delegated sign-in for personal agents. | Let’s say a customer tells their agent to buy a snowboard from your site. With Envoy, the agent signs in under its own identity and places the order within the permissions you set. If the purchase needs sign-off, the customer approves it from their phone. Envoy supports the Personal Agent Protocol from Meta and Sierra and runs alongside your existing user login. | Companies like GoodRx, GoFundMe, and Linktree are already working with Envoy. | |
|
| You read. We listen. Let us know what you think by replying to this email. | | Jensen Huang is out at Nvidia | Nvidia said Thursday that Jensen Huang resigned as president, CEO and board member, effective immediately. CFO Colette Kress, in that job since September 2013, becomes interim CEO while keeping the CFO role. Harvey C. Jones, a director since 1993 and a former CEO and executive chairman of Synopsys, was named the company's first-ever board chair. The search for a permanent CEO covers internal and external candidates. | Huang acknowledged an extramarital affair on Sunday, and Nvidia simultaneously announced an investigation into reports that he had been targeted by a foreign intelligence operation, saying it had "no reason to believe that either national security or NVIDIA secrets were compromised." Fortune reports the woman involved was assessed by US intelligence officials to be a Chinese agent, and that the New York Times reported a second relationship Wednesday, with Nvidia health care VP Kimberly Powell. None of the intelligence allegations have been proven. We covered the original admission Monday and the shareholder suit Wednesday. | NVDA closed down about 4.2% while AMD rose 5.9% and Broadcom 5.5%, which is the market pricing a weaker Nvidia rather than a weaker AI trade. Huang is 63, co-founded the company, and holds roughly 3.6% of it. Kress inherits a business that did $96 billion in first-half fiscal 2027 revenue, up 85% year over year, and she reports Q3 results on Tuesday, November 17. That call just became the most-watched hour in the industry. | Read more. | | Gemini 4 Hydrogen is the one most people will actually use | Google shipped Gemini 4 Hydrogen on Thursday and called it the "daily driver for agentic workflows," leaving Gemini 4 Carbon and Argon to handle deep reasoning. It has four thinking levels, including a new xhigh tier above the default, a 1 million token input window and up to 1 million output tokens, and it acts as the orchestrator in Google Antigravity, switching between Carbon and itself inside a session. | The numbers Google published put Hydrogen at 51 on the Artificial Analysis Intelligence Index at high thinking, above Argon's 50, and 54 at xhigh. On Terminal-Bench 4.1 it hits 57% against 18.2% for Gemini 3.8 Flash, and it finishes tasks roughly 2.5x faster than 3.8 Flash at the default setting. Also on the board: 58.8% on HLE-Diamond, 52.3% and first place on Vals Tax Agent Bench, 55.8% and a tie for first on Vals Legal Research, and 71.6% on Agents' Last Exam v1.1. These are Google's own reported figures. | Pricing is where it gets interesting: $1.50 per million input tokens and $7.50 output, flat across the entire 1M context. Carbon went generally available Tuesday at an introductory $2 and $10 that rises to $4 and $20 on November 30. Hydrogen is rolling out now across the Gemini API, AI Studio, Android Studio, Antigravity for paid Pro and Ultra subscribers, the Gemini app dropdown and Google Search. | Read more. | | Four nonprofits are asking a California court to freeze frontier AI | A coalition including Public Citizen, the Heat Initiative and the Irreplaceable Education Fund sued in San Francisco on Thursday, naming OpenAI, Anthropic, Google, Meta and SpaceX's AI unit. The relief they want is an order blocking development of models more capable than what these companies already ship, which would reach ChatGPT, Gemini and Claude. | The legal theories are California's unfair business practices law and public nuisance, the latter arguing harm to public health from past incidents plus the risk of future ones. Meta declined to comment and the other defendants did not immediately respond. | Courts rarely enjoin an entire industry's R&D, and this is a long shot. What makes it worth watching is the evidence: the complaint leans on the labs' own published warnings. Anthropic's prospectus says internal estimates suggest it "may imminently meet" the acceleration threshold in its Responsible Scaling Policy. Safety disclosure is now plaintiff exhibit material, and that changes what gets written down. | Read more. | | | | Google also shipped the Gemini Decisions API, which is the most practical thing it has released this month. Instead of generating text, the gemini-decision-1 model scores one shared input against up to 1,000 yes/no, multiple-choice or rubric questions in a single pass, returns a confidence value for each, and runs in roughly 150 to 250 ms. It is $0.10 per million input tokens with output free, and the shared input bills once no matter how many questions ride along. For routing, policy checks and content grading, that is a different cost structure than wrapping a chat model. | On the security side, OpenAI launched Daybreak Red Team and Defense Factory, which probe applications for weaknesses and propose patches, with Red Team in early access and human engineers reviewing fixes before deploy. Anthropic opened a private beta for self-hosted Claude Managed Agents, letting companies keep agent execution and session history on their own servers. And IBM previewed Granite 4.1 LLMON in 3B and 8B, trained to separate instructions from untrusted data using explicit markup, which is a more interesting answer to prompt injection than another filter. | In models and hardware: Thinking Machines put Inkling 2 Small on Hugging Face under Apache 2.0, Qwen opened Qwen4-Max-Preview through QwenCloud's Token Plan until November 25, World Labs opened Atlas to everyone for reconstructing 3D scenes from images, and Google and Samsung opened preorders for Gemini smart glasses from Warby Parker and Gentle Monster starting at $395, shipping December 10. AWS also released a free local emulator covering 200-plus services and says AI agents wrote and maintain all of its code with no manual engineering review, which is either the most impressive or most alarming line in today's edition. | | Thank you for reading today’s edition. | | Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter. | Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X! | Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months. |
|