Huang resigns at Nvidia; Gemini 4 Hydrogen at $1.50/M; 35,578 Grok Build apps open to anyone

🔥What's hot

🛠New tools & products

• Self-hosted Claude Managed Agents — a private beta that runs Anthropic's whole agent service inside your own cloud or data centre: harness, sandboxes, credentials and session history. This is the answer to the regulated client who says agent runs cannot leave the network, and it kills the build-it-yourself platform project. Credentials sit in vaults outside the sandbox and a proxy injects them into MCP tool calls; session records export over OpenTelemetry into monitoring you already run. Claude by Anthropic

• Claude Sheets and Whiteboards — spreadsheets and shared brainstorming boards that Claude builds with you inside the chat, plus the renamed Claude Data dashboards, now generally available on paid plans. Expect the "does this replace Office?" question; Anthropic's own product lead calls Sheets fairly basic and points advanced work back at Microsoft and Google. Claude by Anthropic

• MCP file transfer extension — a draft addition to the Model Context Protocol, the standard wiring agents to tools, giving clients and servers one agreed way to move files up and down. Two years of per-project hacks, from temporary S3 links to bolt-on upload pages, end if this lands. Files bypass the model: small ones ride inside the tool request, larger ones upload separately. Maintainers target a stable release by year-end, and support stays optional. The New Stack

• J.Crew Style Advisor — an AI stylist on JCrew.com that answers "what do I wear to a beach wedding" with complete outfits photographed by the brand's own stylists, built with Claude Code. A clean client-facing example of model tiering: a frontier model designs the behaviour, cheaper models run it in production. Fable 5.1 planned the stylist reasoning, Opus 5.5 built the prototype, and each live conversation uses Sonnet 5.5 as the stylist with Haiku 5.5 on quick tasks. Claude by Anthropic

• Agents get shared boards, and authors review their own PRs. A roundup of this week's delivery patterns: Google Swarm and Cursor Projects give parallel coding agents shared boards and direct messaging, and GitHub pushes reusable, versioned skills that carry team conventions between agents. This is the practice layer clients reach for once the model question is settled. The same roundup notes Amplitude's Wave agent compressing product fixes to one or two days, and a team that swapped peer review of AI-written changes for author-led review backed by CI. o16g

💬What people are saying

📰News

• Gemini 4 Hydrogen — Google's cheap, fast Gemini 4 variant, pitched as the executor that writes code and runs tests while Carbon and Argon do the hard thinking. The price frames every cost conversation: $1.50 per million input tokens and $7.50 output, flat across a million-token context, which Simon Willison notes is 15 times what Claude Haiku 5.5 charges. Antigravity's new Orchestrator swaps models mid-session; on Terminal-Bench 4.0 it neared Carbon's success rate on 24% less quota, finishing 19% faster. Google

• Anthropic's IPO — the roadshow opened on 10 November with 825 million Class A shares at an indicative $70 to $78, listing on Nasdaq as ANT. Clients will ask for the number: up to $1.4 trillion fully diluted, below the $2 trillion Reuters says was once expected. The company's IPO page says pricing is expected on 18 November with trading due to begin 19 November; filings show roughly $81 billion annualised revenue against nearly $900 billion of multi-year compute commitments. Anthropic

• CodeRabbit — a review service that checks pull requests independently of whichever agent or human wrote them, and its subscription revenue has now reached $100 million a year, up from $10 million a little over a year ago. The loudest money signal yet that clients pay for verification, not generation. It claims over 20,000 customers and three million pull requests reviewed weekly, and cites June 2026 New Relic research where 62% of tech leaders said teams often merge AI-written code nobody read closely. CodeRabbit

• OpenShell — Nvidia's open-source runtime that pens an agent inside a sandbox and enforces which files, networks and credentials it can reach; Nvidia has now handed control to the Cloud Native Computing Foundation. Neutral governance is what procurement wants to hear when you propose where agents will run. It enters at the CNCF's entry-level Sandbox tier, uses Cedar as its policy language, and its project page lists about 380 contributing organisations. Techstrong.ai

• Grok Build — SpaceXSI's chat-to-app builder gives every app its own web address, and security firm Claroc says almost every live one opens without a sign-in. If a client lets staff ship internal apps from a chat box, this is the default setting to go and check this week. Claroc loaded 37,571 live addresses and 35,578 opened without sign-in, including 24 built by SpaceXSI and X staff; the random words sit in the hostname, so passive DNS lists them. Ars Technica

• Agents "found" a Maya site that was already published. A Google researcher set 48 agents loose on NASA lidar scans of Mexican jungle; they flagged an unregistered hilltop with a plaza and a pyramid, which an update to his post admits archaeologists had published in 2025. The retellable lesson: the agents read the terrain fine and fluffed the literature check nobody had named as the real test. Mexico's heritage institute confirmed no registered site at the spot, which is not the same as unknown; a coauthor of the 2025 paper worries about a flood of AI discovery claims. pietromiotti.com

💡One takeaway