| Weekly Edition |
Price: Zero tokens |
|
| The Weekly Token |
| All the week’s news that fits the context window. |
|
| Vol. I · No. 5 · November 7 – 14, 2026 |
|
|
Hi,
Welcome to issue five of The Weekly Token. As always, these are the week’s biggest AI announcements, curated from X, with a link to every original post so you can go straight to the source.
This was one of the busiest weeks I’ve tracked. Jensen Huang resigned as NVIDIA’s CEO after a week-long board investigation. Anthropic shipped Claude Fable 5.2 and Google answered with Gemini 4 Carbon, Hydrogen and a new Decision model. Anthropic also filed for its IPO, and ChatGPT’s dots got a wallet. My picks still lean towards what matters if you’re building with agents.
The Agentic UI section is packed this week. Google shipped A2UI 1.0 as a stable open protocol for generative UI, and Claude now runs full apps beside the chat through more of the MCP Apps extension. Both are worth a close read if you build interfaces for agents.
As always, hit reply and tell me what was useful, what was noise, or what I missed.
|
|
|
Lead story · Industry
NVIDIA's board said on Sunday that it had opened an independent investigation into reports that Jensen Huang was targeted by a foreign intelligence operation, while stating that it had confidence in his leadership. By Wednesday, Andrew Curran was relaying a Wall Street Journal report that the board was preparing to oust him and planned to meet later that day.
On Thursday NVIDIA's newsroom announced that Huang had resigned as president, chief executive officer and a member of the board, effective immediately. Colette Kress, CFO since 2013, will serve as interim CEO while remaining CFO, and the board has begun a search that will consider internal and external candidates. Harvey C. Jones, a former CEO and executive chairman of Synopsys, became the company's first-ever Chair of the Board and said operations, customer commitments and the product roadmap would continue without interruption. The post drew more than 72,000 likes, and tributes followed from across the industry, including Perplexity CEO Aravind Srinivas.
|
|
|
A2UI 1.0, released under Apache 2.0, lets agents answer with native components from an app's own catalog instead of streaming raw HTML. Google's launch article says UI messages are pure declarative data, adds Express, a compact DSL inspired by OpenUI Lang that cut generated UI tokens by 63% in Google's benchmark, ships Python, TypeScript and Dart agent SDKs with renderers from React to SwiftUI and Jetpack Compose, and embeds MCP Apps directly for widgets that need an iframe. CopilotKit, which co-contributed to the launch, followed on Thursday with Open Intelligent UI, built on A2UI 1.0, and OpenUI's Rabi Shanker Guha said OpenUI and A2UI interop is coming soon.
|
|
|
Claude can now open the full version of a tool inside the conversation, starting with Adobe, Atlassian and Box, rolling out on web and desktop, so you can make changes yourself or ask Claude to make them with you. Anthropic's developer account said Claude now supports more of the MCP Apps extension, including app tools, which Claude calls to change what an app shows, and embedded sites, so a full app can run inside Claude. Box said it is a launch partner.
|
|
|
Vercel announced the UI SDK with a launch video. Guillermo Rauch framed it as drilling deeper than component libraries for the agents era. Posts describing the beta repo list keyboard, focus, selection and ARIA behaviors with framework adapters.
|
|
|
Michael Kolkov shipped a sports-themed iOS demo that shapes the interface from a request instead of pushing frequent app updates. He says abstract, catch-all fields are too full of edge cases, but narrowing to a specific theme unlocks smarter UI than a set of conventional primitives. The project includes a preview studio, a component registry, and documentation for each layer.
|
|
|
CopilotKit's Atai Barkai lists generative UI, computer use and routines; it is on Mac and Windows.
|
|
|
Tools registered with App.registerTool() are mirrored into document.modelContext, and page tools are served back to the host.
|
|
|
It hit No. 2 on GitHub's trending repos the same day.
|
|
|
|
Organizations can host the agent harness, sandboxes and session history in their own cloud or on-premises, and choose where requests to Claude go. Anthropic pitches it for work that needs internal code and systems, like fixing broken builds or triaging alerts. Eligible orgs request access through their account team.
|
|
|
The new /swarm slash command turns one prompt into a multi-agent workflow in the desktop app and CLI, with agents planning, building and verifying together on a live message board, and each role can use a different model. A day later Antigravity rolled out browser use, so its agents can research, write code and verify it live in a browser; users choose which sites to log into, and downloads pause until approved.
|
|
|
OpenCode's Browse lets an agent control the browser you already use, with its open tabs, saved logins and extensions, rather than an empty session. The product page describes it as a way for OpenCode and other MCP clients to drive background tabs with Playwright, handing control back to you when a site asks for a login or two-factor code.
|
|
|
Guillermo Rauch said MCP could read and write but could not report later that an event happened, and that MCP Events fixes this. Vercel is adding support to its own MCP server, so an agent that starts a deployment can subscribe to deployment.succeeded.
|
|
|
|
Anthropic released Claude Fable 5.2 and the more restricted Claude Mythos 5.2 on Monday, citing gains in coding, reasoning and vision. Fable 5.2 is up to 30% faster than 5.1 and costs $8/$40 per million input/output tokens, down from $10/$50, and counts for up to 50% of Claude plan limits until Nov. 16. Artificial Analysis put it No. 1 on its Intelligence Index at 65, Cursor said it tops CursorBench at 60% at max effort, and it led DeepMind's new MKA benchmark at 57% and AA-Robotics at 75. Not every test went its way: on Friday Deepnote's new Agent Initiative Benchmark, which drops models into a synthetic company's Slack, ranked OpenAI's GPT-6.1 Sol and GPT-6 Astra ahead of it, with Sol at $0.11 per task against Fable's $1.87.
|
|
|
On Tuesday Google released Gemini 4 Carbon, its strongest frontier model, built for long-horizon work in software engineering, cybersecurity, knowledge work and science; Cursor said it tops CursorBench at 60.4% at high effort, and Perplexity said it matches Fable 5.2 on its WANDR benchmark at 40% of the cost. Carbon is for Ultra subscribers in the Gemini app and through the API. On Thursday came Gemini 4 Hydrogen, which DeepMind says finishes tasks 2.5x faster than Gemini 3.8 Flash and beats Gemini 4 Argon on the Artificial Analysis index; it is rolling out to all Gemini app users and is priced at $1.50/$7.50 per million tokens, per Cursor. Sundar Pichai's Friday recap also listed two new TPUs, Gemini Robotics 2.5 and AlphaGenome 2.
|
|
|
The new Gemini Decisions API scores up to 1,000 closed-set questions across text, images, audio, video and PDFs in one pass, with yes/no, multiple-choice (up to 255 options) and rubric-scoring modes. It costs $0.10 per million input tokens with nothing billed for output, which Pichai put at about five times cheaper than Flash-Lite, and Google pitches it for routing, tool selection, safety checks and grading inside agents. Logan Kilpatrick said it has a 256K context window and is free to try in AI Studio.
|
|
|
Meta says Muse Ember is built for token-heavy agent work such as tool calls, coding subtasks, routing and extraction, and has a 1M-token context window. Chief AI Officer Alexandr Wang said it delivers more intelligence per dollar than Haiku, Luna and other flash models at a third of the cost or less. It is on the Meta Model API and Muse Code, and later in the day Meta brought Muse Ember, Muse Spark 1.3 and Muse Wonder to Google Cloud.
|
|
|
The Qwen team says Qwen4 brings deeper knowledge and stronger coding and agentic capabilities. Qwen4-Max is in preview on Alibaba's Token Plan, with early access in QwenWork, Qoder, the Qwen app and QwenStudio.
|
|
|
|
On Tuesday Anthropic began rolling out three new Claude surfaces: Whiteboards for thinking a plan through, Sheets for building trackers and Data for telling a story with your data. On Thursday came Claude Meetings, which joins a video call as a visible participant, builds docs, whiteboards and other artifacts while people talk, and can be asked questions out loud or in chat; notes, action items and transcripts land on the meeting's page afterward. It starts with Google Meet in beta on Team and Enterprise plans. Anthropic also gave paid users one free full limit reset, usable before Dec. 10.
|
|
|
ChatGPT Wallet lets a dot complete purchases on your behalf without access to your actual card details, and every purchase requires a security check. It is rolling out to dots for Pro subscribers in the US, and Greg Brockman said it was built with American Express, Mastercard and Visa. It capped a busy week for dots: on Monday OpenAI let users connect several local computers to the same dot, and on Thursday dots got faster replies and began following custom instructions.
|
|
|
Search, files, links, safety and voice, the capabilities behind the DeepSeek App, are now available through the DeepSeek API, switched on with a few parameters, alongside standalone OCR, ASR and TTS APIs. DeepSeek also released an open-source frontend demo that can import DeepSeek conversation history. Earlier in the week, Bloomberg reported that DeepSeek Harness, its open agent platform, had reached 1.54 million daily users.
|
|
|
|
NVIDIA calls GR00T 2 its most advanced open System One model for robotics research. It learns skills from human demonstrations and predicts how its actions will change the physical world, and those skills can carry across different robot bodies, reducing the need to teach each robot every task from scratch. It arrived the day after Jensen Huang's resignation.
|
|
|
The new model predicts how a DNA variant affects specific cell types, helping explain why one single-letter change can drive disease in some cells and spare others. Demis Hassabis said it is available for academic research through an API.
|
|
|
New OpenAI research tracks no-CoT capability and CoT controllability across the training lifecycle. The authors call the finding "unsurprising but sobering": as models scale, monitoring them through their chain of thought gets harder.
|
|
|
|
The weights are now on Hugging Face. Together AI describes it as a multimodal MoE for coding, long-horizon agentic work and vision, with 594B total parameters, 17B active and up to 1M tokens of context. LMArena says it is the top US open model in Code Arena: WebDev.
|
|
|
Georgi Gerganov released llama.cpp v1.0.0. Hugging Face CEO Clem Delangue, announcing it with Gerganov, called it the first major fully stable version of the most used library for running AI models locally on phones, laptops and workstations, and said it shows a commitment to support every hardware, platform and model type.
|
|
|
On Thursday Anthropic said it has filed a registration statement, including a prospectus, with the SEC for its initial public offering, and pointed investors to an offering site and the filings on EDGAR. On Friday it streamed a fireside chat between Dario Amodei and JPMorgan CEO Jamie Dimon. In a report shared on Monday, Andrew Curran said the shares could start trading the following Thursday or Friday.
|
|
|
At its "Cyber at the Frontier" keynote on Thursday, OpenAI announced two cyber products, which Greg Brockman said make it much easier to put frontier cyber capabilities to work defending companies. Thibault Sottiaux described them as tools for automated patching and automated red teaming and posted the recording. The keynote followed last week's GPT-6 Cyber launch.
|
|
|
Stripe said OpenRouter is now part of the company and that together they will help businesses explore new models and optimize token routing and usage. TestingCatalog quoted OpenRouter saying its commitment to neutrality stays the same.
|
|
|
| ■ |
Google DeepMind's Gemini Robotics 2.5 learns hundreds of everyday skills — From folding laundry and serving pastries to sorting car parts on a factory floor.
|
| ■ |
DeepMind's AlphaProof Code Proves Signal's Protocol Secure Down to the Code — Google DeepMind says that for the first time AI has mathematically verified the Signal messaging app's cryptographic system at the code level, using its AlphaProof Code framework.
|
| ■ |
Claude Code Projects get a more proactive coordinator across parallel threads — Custom desktop themes, called mods, arrived the same day.
|
| ■ |
Codex Lead Clarifies Pro Plans: No 5-Hour Limits, Up to 50x Plus Usage — OpenAI's Tibo Sottiaux said all Pro tiers drop the 5-hour limit: Pro 100 gives 10x Plus usage, Pro 200 gives 20x, and Pro 500 gives 50x plus access to "ultrafast." In a follow-up he asked users whether a Pro 50 tier would make sense.
|
| ■ |
OpenAI Launches Businesses in ChatGPT in Beta — Businesses can claim their listing, manage their information and see how people engage with them. Eligible businesses can also build a Business Agent to answer customers' questions.
|
| ■ |
ChatGPT Finances Comes to the Desktop App — Users can find ways to save, understand spending, make sense of investments and build custom financial tools. It is available to Free, Go, Plus and Pro users in the U.S.
|
| ■ |
Edge0 releases Edge0-145B and Edge0-120B, open models built to run on a laptop — The team says the 145B model needs about 29 GB of memory and scores 86.5% on Terminal-Bench 2.1.
|
| ■ |
MiniMax Releases H3-Image With Open Weights — MiniMax says H3-Image generates 16-megapixel images in about 10 seconds, follows prompts precisely, edits, and takes up to 16 reference images. It is on Hugging Face, the MiniMax API, MiniMax Code and Hailuo AI, with 50% off 16MP+ generations during launch week.
|
| ■ |
Baidu Previews ERNIE 5.5 Flash, Open Weights Due by Nov. 25 — The open-weight model has 280B total parameters with 14B active per token, a 1M-token context window and generation above 200 tokens per second, aimed at coding, 3D modeling and knowledge work. Arena says it is trending toward No. 2 among open models on Code Arena WebDev at 1668 points.
|
| ■ |
Runway Releases Praxis-1 Robot Model With Open Weights — Runway made its world action model available to everyone. It turns video pretraining into control for real robots and ships under Apache 2.0, ready to run and fine-tune on your own hardware.
|
| ■ |
World Labs opens its Atlas world model to everyone — Apps and API endpoints turn ideas into keyframes, video clips or 3D Gaussian splats.
|
| ■ |
OpenAI Details Control Room, a System to Halt Research Workloads in 35 Seconds — OpenAI says Control Room tracks every research inference workload, including agentic sampling for training and evaluation, and works with its misalignment monitoring to halt model output within 35 seconds of a decision to stop or pause a job.
|
| ■ |
MaxxBench Finds Popular Coding Scores Fall Apart After Small Task Tweaks — Small tweaks to 89 DeepSWE tasks spread models that clustered near 74% by 23 points.
|
| ■ |
Anthropic publishes 10 recommendations for designing, running and reporting safety evals |
| ■ |
Anthropic Explains How Claude Agents "Dream" to Repair Their Own Memory |
| ■ |
n8n Labs Introduces Nod, an Open-Source Personal Agent — Nod runs on your own computer with its own browser to prep calls, collect invoices or cancel subscriptions across apps and websites, using the AI provider you choose.
|
| ■ |
Cursor Rebuilds Projects for Multiplayer — Teammates and their agents can now work together in one shared workspace, and agents can coordinate with teammates' agents.
|
| ■ |
Cursor publishes an update on its recent security incident and remediation steps |
| ■ |
Google Adds TPU 8i Max and TPU 8tx, Completing a Four-System TPU Generation — TPU 8tx for confidential computing and TPU 8i Max for huge context windows complete a four-system generation.
|
| ■ |
Report: OpenAI's Q3 Revenue Hit $10 Billion — Andrew Curran cites The Information, which credits strong demand for Astra.
|
| ■ |
OpenAI Acquires Travel Startup Odessia — Odessia's founder says the team will focus on the future of travel at OpenAI.
|
| ■ |
ChatGPT Retires Advanced and Standard Voice on Dec. 11 — OpenAI says nearly all voice users are on GPT-Live; live video and screen-sharing keep Advanced voice for now.
|
| ■ |
Google DeepMind releases PARDIG and SatWeave, the datasets behind WeatherNext 3 |
| ■ |
Nous Research Releases Arete-27B, a Western-Values Post-Train of Qwen3.8-27B — After fine-tuning on about 53 million tokens of SFT, Arete-27B's rate of China party-line answers on an 803-prompt values eval fell from 22% to 1%, in line with Western models Nous tested, with no meaningful drop on Hermes Bench. Eval, weights and training data are public.
|
|
|
|
That’s the week. If one of these changes how you build, or you think I picked the wrong stories, reply and tell me.
See you next Saturday, Apurv
P.S. I also break this stuff down on YouTube and at apurvkhare.com.
|
|
Sourced from X only · every headline links to its post The Weekly Token · Vol. I · No. 5 · Week of November 7 – 14, 2026 |
|