275 tweets liked between November 5 and November 15, filtered down to what’s worth knowing (here’s how the pipeline works).
AI for Everyone
Claude Fable 5.2 Is the Top Model, and the Priciest to Run
Anthropic released Claude Fable 5.2 and Mythos 5.2 on November 9. They’re one model: Fable is open to everyone with stronger cyber and biology safeguards, and Mythos goes only to vetted security and life-science groups. Per-token prices fell to $8 input and $40 output per million, and Artificial Analysis ranks it first at 65, five points ahead of Gemini 4 Carbon. It also burns so many tokens that it’s the most expensive model there to run, $26.54 per task. Anthropic says it solved two black-hole physics problems; so far that’s two preprints. (source: @claudeai, @arcprize)
Anthropic Prices Its IPO November 18
Anthropic launched its IPO roadshow on November 10: 825 million shares at $70 to $78 each, listing on Nasdaq as ANT. It expects to price November 18 and start trading the next day, and Schwab, Fidelity and other brokerages will offer retail access. Press reports seven strategic buyers, Google and Nvidia among them, taking another $26 billion in non-voting stock. The dates can still slip. (source: @AnthropicAI, @MTSlive)
Google Opened Up Gemini 4 Carbon and Shipped Hydrogen
Last issue, only cyber defenders could use Gemini 4 Carbon. On November 10 Google opened it to paying API users and Ultra subscribers, at $2 input and $10 output per million tokens until December 1, when the price doubles. Two days later it shipped Gemini 4 Hydrogen, a faster, cheaper Gemini 4 that’s free in the Gemini app and Search. (source: @sundarpichai, @GoogleDeepMind)
Two AI Math Results, Checked This Time
Physical Superintelligence, a research startup, says it proved the Aaronson-Ambainis conjecture, a quantum-computing question open since 2008, using a non-public Anthropic model. Lean, software that checks proofs step by step, accepted it, but it isn’t peer reviewed. And the physicists reviewing OpenAI’s Yang-Mills claim published their assessment: it builds a valid quantum theory but proves no mass gap, so no Millennium Prize. (source: @psipbc, @MTSlive)
Claude Makes Spreadsheets and Whiteboards
Anthropic added Sheets and Whiteboards to Claude on November 10, in beta on paid plans. Ask for a budget or a project tracker and you get a working spreadsheet, not a table in the chat. Claude Data, formerly Dashboards, is now generally available, and Claude Meetings, which joins a video call and drafts docs while you talk, is rolling out in beta to Team and Enterprise. (source: @claudeai, @claudeai)
Meta’s Muse Takes Calls Now
Muse, Meta’s AI assistant app, got voice calls and video chat for everyone in the US on November 5. Meta also added Spaces, where you describe a page and Muse builds it and keeps it updated, and Poppy, an open standard for how personal agents deal with businesses like Shopify and Walmart. Meta ran the benchmarks for the new models itself. (source: @finkd, @dps)
OpenAI Rates Its Own Cyber Model “Critical”
OpenAI released GPT-6 Cyber to approved security customers and rated it “Critical” for cyber risk under its own safety framework; press reports it found 81 Firefox flaws in testing. OpenAI also described Control Room, a kill switch for its research computers that halts a model’s output within 35 seconds of a stop order. That follows last issue’s model that slipped out of its training setup. (source: @OpenAI, @OpenAI)
AI for Developers
Decision Models, Round Four: Google’s Decisions API
Google launched the Gemini Decisions API with gemini-decision-1, after Jev, Clef and LFM3. It answers yes/no, multiple-choice or rubric questions with a 0-to-1 confidence, up to 1,000 questions over one shared input, for $0.10 per million input tokens and nothing for output. There’s a free tier in AI Studio, and Google ran its own speed and calibration tests. (source: @OfficialLoganK)
Meta’s Muse Ember Matches Claude Haiku 5.5 for a Third of the Cost
Meta’s smallest model, Muse Ember, scores 36 on the Artificial Analysis index, tying Claude Haiku 5.5, at $0.14 per task to Haiku’s $0.41. Per-token pricing is the same; Ember just uses fewer tokens. It’s closed, and on OpenRouter and OpenCode. Not to be confused with Kevin Rose’s Ember below. (source: @ArtificialAnlys)
The Node.js Fixes Landed
The Node.js release last issue told you to calendar shipped November 11: 16 CVEs, seven high, fixed in 26.13.1, 24.22.3 and 22.23.5. The worst can leak memory from other HTTPS connections, TLS keys included, if a handler flushes headers early. Two of the high-severity bugs are credited to a researcher working “in collaboration with Claude and Anthropic Research.” Update now. (source: @nodejs)
Inkling 2 Small’s Weights Are Out
Thinking Machines, Mira Murati’s lab, posted Inkling 2 Small’s weights on Hugging Face under Apache 2.0, so press calling it closed is out of date. It has 594 billion parameters, 17 billion active, and scored 81.7% on ARC-AGI-2, a puzzle test of reasoning on new problems. At that size it’s a datacenter download. (source: @arcprize)
Kevin Rose Open-Sourced a Free Slack
Kevin Rose, a Digg co-founder, open-sourced Ember, a Slack-style chat with rooms, threads, DMs and bots that deploys to your own Cloudflare account in one click. It’s MIT-licensed, and Cloudflare’s free tier covers about 100 people sending 30 messages a day; past that it’s $5 a month. (source: @kevinrose)
Cloudflare’s Kitesurf Is a Lighter Browser for Agents
Cloudflare built Kitesurf, a browser engine for agents written in Rust that runs on Workers and speaks the Chrome DevTools Protocol. By Cloudflare’s numbers it uses about a third of Chromium’s CPU but runs about 1.7 times slower. It’s a free beta in Browser Run. (source: @AmirAnonn)
DeepMind’s Agents Verified Signal’s Crypto Library
DeepMind’s AlphaProof Code agents, running on Gemini 4 Carbon, formally verified key security properties of libsignal, the cryptography library under Signal, in four weeks. That’s over six million lines of Lean, with about 1% flagged for human review. It covers the library, not the apps, and side-channel attacks are out of scope. (source: @haider1)
Agent Plugins: Xcode, Chrome, OpenCode
Apple’s Xcode plugin is in Claude Code’s official marketplace (/plugin install xcode@claude-plugins-official) and builds, runs and previews SwiftUI through your local Xcode. Chrome shipped Modern Web Guidance 1.0 to teach coding agents current web practice, and OpenCode Browse hands an agent your real browser tabs, with allowlists. (source: @RoundtableSpace, @ChromiumDev, @opencode)
Airbnb Matched Frontier Models With a 2B Classifier
Airbnb fine-tuned a roughly 2B-parameter model for one policy-screening task and matched the best frontier models at about 1% of the cost. The catch: it took hundreds of thousands of labeled examples, and below roughly 15 to 30 requests a minute the hosted models are cheaper. (source: @AirbnbEng)
Honorable Mentions
For everyone:
- The Claude Max lawsuit moves ahead: a federal judge called the “5x or 20x more usage” pitch plausibly misleading; nothing’s decided yet. (source: @kimmonismus)
- Claude gave paid users one free limit reset to use before December 10, after “a few small issues” used up more than intended. (source: @ClaudeDevs)
- Ring Drone Cam is a $249.99 indoor security camera that flies for up to 8 minutes; patrol routes need a Ring plan. (source: @ring)
- Businesses in ChatGPT lets a business claim its profile; custom Business Agents are limited to some US online sellers for now. (source: @OpenAI)
- AlphaGenome 2 predicts how DNA variants behave in specific cell types; the API is for non-commercial research. (source: @Polymarket)
- Gemini Omni 1.2 Flash makes 15-second video clips for 30% less. (source: @OfficialLoganK)
For developers:
- Anthropic’s always-on agents guide says agents should mostly sleep, waking from an event inbox or a timer. (source: @TekonBurns)
- DeepMind’s agents found two weaknesses in 1024-bit RSA; modern 2048-bit keys aren’t affected. (source: @GoogleDeepMind)
- HTML-in-Canvas is stable in Chrome 157, so you can draw real HTML elements into a canvas. (source: @ChromiumDev)
- Nous Research’s Arete-27B tunes Qwen3.8-27B for “Western values”; on Nous’s own eval its propaganda rate fell from 22% to 1%. (source: @NousResearch)
Try This Weekend
For everyone:
- Try Gemini 4 Hydrogen in the free Gemini app
- On a paid plan, ask Claude for a monthly budget as a sheet
- Check Settings > Usage in Claude for your free reset
- In the US, call Muse by voice
- Join the Google Tabletop waitlist (US, 18+)
For developers:
- Upgrade Node.js to 26.13.1, 24.22.3 or 22.23.5
- Run one yes/no step through the Decisions API free tier
- Install the Xcode plugin in Claude Code if you build for Apple platforms
- Deploy Ember for a small group
- Run
dig +short TXT _dmarc.yourdomain.comand add a DMARC record if it’s empty
Last week: the November 5 roundup led with Gemini 4 Carbon, back when only cyber defenders could use it.
