Issue November 19Daily digest · 13 items
In this issueReviewed 35 sources, gathered 13 items
Studio still life of two cards with a gold $7.50 receipt and a heavy timer
ArticleAILLMAnalyticsProductivity

GPT-6.2 Astra vs. Claude Fable 5.2 Across 12 Tasks: 7.2x Cheaper, So Why Are Testers Arguing?

In a head-to-head test on real business projects, GPT-6.2 Astra and Claude Fable 5.2 acted like two completely different specialists. Astra bypassed a Canva stroke limit by moving the canvas and finished in 90 minutes for $7.50, while its rival spent $63 over three hours and gave up. Yet on a $15 Excel financial model, a 3D physics game, and a five-hour video edit, Fable showed the depth that testers say OpenAI's new release is missing.

Choosing between these models is no longer a simple race for the smartest AI on a leaderboard.

👍 23
Nate HerkNick SaraevJulian Goldie SEOOpenAI API Pricing — GPT-6.2 Astra and GPT-6.2 Sol+5
November 19
Patent-style blueprint of a recursive model self-improvement loop with fine lines and a risk gauge
NoteAISafetyLLMAI

GPT-6.2 Astra Hits "High" Rating for AI Self-Improvement: What the Model Can Actually Do

Dramatic headlines claim AI is already rewriting itself, yet OpenAI's report points to practical tasks like finding bugs and speeding up compute under that label. Meanwhile, labs are splitting: some release these capabilities openly, while others lock down against rivals with strict filters and barriers. Even choosing how long new models think works against intuition, and maxing out the slider easily wastes your limits.

👍 10
David OndrejOpenAI — GPT-6.2 Astra and GPT-6.2 Sol System CardEpoch AI — Epoch Capabilities Index (ECI)Anthropic — Detecting and preventing distillation attacks+1
November 19
Isometric diorama of a digital office with a task board, spreadsheet, and charts
ArticleAIProductivityAutomationAnalytics

An Office Inside Claude: Whiteboards, Formula-Driven Sheets, and Videos Built From Data

A single Claude chat now brings together a brainstorming board, a formula spreadsheet, and video reports built on live databases:

  • Whiteboards turns a spoken plan dictated in 25 languages into a timeline and task cards.
  • Sheets moves those tasks into a live spreadsheet with formulas, shared edits, and Excel export.
  • Data connects to Ahrefs or BigQuery and turns raw numbers into a scrollytelling page and a one-minute video.

The real catch starts when your prototype hits the ceiling of pivot tables and company compliance rules.

👍 20
Skill Leap AIIntroducing Claude Whiteboards, Claude Sheets, and Claude Data | Claude by AnthropicGet started with Claude Whiteboards | Claude Help Center
November 19
News · November 191 news item

Outages at Microsoft, US Aviation, and Chase Bank: Where AI Fits In and Why It Wasn't a Cyberattack

Rumors of an "AI cyberattack" hitting Microsoft, US flights, and JPMorgan Chase fell apart against ordinary causes: a network update, a single router in Arkansas, and a software bug. Yet AI is still part of the story—banks are rushing to patch old code after AI audits like Mythos, and that speed raises the risk of new outages.

Workspace with an open video call window and documents syncing in real time
NoteAIAI AgentsProductivityCommunication

Claude Meetings in Google Meet: How AI Builds Docs and Whiteboards Live on Your Call

Why did Anthropic put Claude directly into Google Meet when regular AI bots already record calls?

The model builds documents, slides, and Claude Whiteboards live during the conversation, and teammates edit them right away in real time. Personal accounts and Zoom users are locked out of the beta for now: it only runs on Team and Enterprise plans with an eight-hour call limit. On a private meeting, even a single third-party notetaker in the room will make the assistant leave immediately.

👍 9
Julian Goldie SEOIntroducing Claude Meetings, now in beta | Claude by AnthropicUse Claude Meetings | Claude Help CenterAnthropic launches Claude Meetings to turn workplace discussions into plans and tasks | TechCrunch
November 19
Minimalist typographic poster showing a stack of paper notes turning into a website
NoteAIAIProductivityData

Gemini Notebook Now Builds Interactive Sites From Notes and Tables in One Prompt

An interactive site from one sentence sounds like just another landing page builder, yet Gemini Notebook refuses to invent data. This free web tool builds clickable dashboards, trackers, and explainers strictly from your uploaded spreadsheets and notes, complete with links to the original sources.

That convenience comes with strict sharing hygiene: anyone viewing a shared page gets access to the entire notebook, and each viewer only sees their own checkboxes. Clean source files make the difference between a useful tool and a generic placeholder.

👍 12
Julian Goldie SEOTurn Gemini Notebook sources into interactive sites — Google BlogGenerate sites in Gemini Notebook — Gemini Notebook Help
November 19
Exploded engineering diagram of a two-model setup with a heavy planner and a lightweight accelerator
NoteAILLMAI AgentsDevelopment

Gemini 4 Hydrogen: How Google's Fast Agent Executor Works and Why It Teams Up With Carbon

Should you run one heavy model for every step of development, or split planning and fast execution? In the Google Antigravity orchestrator, senior Gemini 4 Carbon builds the plan while Gemini 4 Hydrogen edits code across a 1M-token window at $1.50 input and $7.50 output per million.

The pair uses 24% less quota and finishes 19% faster at flagship quality, while Hydrogen alone hits 57% on Terminal-Bench 4.1. There is one area, though, where you still need to keep a tight rein on it.

👍 4
Julian Goldie SEOIntroducing Gemini 4 Hydrogen for Agentic Workflows — Google BlogGemini 4 Hydrogen — Model Card, Google DeepMind
November 19
Macro photo of a silicon chip with gold micro-traces and high-speed light streaks
NoteAILLMAI AgentsAutomation

Tencent HY4 Flash: 224B Parameters, 150+ Tokens/Sec, and Two Free Weeks in Hermes

In a blind test across 259 work tasks, the new HY4 Flash scored 2.79 vs. 2.81 for Claude Haiku 5.5 and posted the fastest median time at 6.7 minutes. Task cost dropped roughly 18x, though DeepSeek V4.1 Flash still led on user satisfaction in the same comparison.

You can run the model free for two weeks in Hermes Agent with a 1M-token context, just watch out for a paid twin sitting right next to it in the catalog. This setup handles office routine easily, though you shouldn't leave it alone on heavy coding.

👍 6
Julian Goldie SEOTencent Launches Hy4 flash to Power High-Speed, Cost-Efficient AI Deployment for Everyday Workflowstencent/Hy4-flash — Model Card & Tencent Hy License v1.0 (Hugging Face)Nous Portal — Model Catalog & Pricing+1
November 19
Modular clay composition of geometric blocks with precise perforations and data structure
NoteAIAutomationDataLLM

Liquid AI d1.2 and d1.2-ultra: Why Automation Needs Decision Models Instead of Chatbots

Instead of writing long replies, Liquid AI d1.2 decision models return strict JSON in a single pass and charge only for input tokens.

  • Extracting form fields with an honest "not found" instead of guessing.
  • Finding text spans in documents and object coordinates in photos.
  • Counting items with a range and tagging one message with multiple labels.

JSON formatting alone isn't enough for reliable automation: you need an honest confidence score on every answer.

👍 9
Julian Goldie SEOd1.2: State-of-the-Art Omnimodal Decision Models — Liquid AI BlogLiquid AI d1.2-ultra — Artificial Analysis Decision IndexLiquidAI: d1.2-ultra — API Pricing & Benchmarks (OpenRouter)
November 19
Cinematic shot of a server cluster with rhythmic rows of racks and optical lines
NoteAIAI SearchAIAI Agents

Perplexity Frontier Research: Why the Search Engine Dedicated 10,000 GPUs to an Autonomous AI Scientist

A search company that spent years sitting on top of other labs' models has dedicated over 10,000 GPUs and hired NYU professor Andrew Gordon Wilson to build an autonomous AI researcher. The catch with Frontier Research is that continual learning without forgetting remains a future goal.

Its engineering base is already paying off, though: open-source fabric-lib pushes 400 Gbps at 97.1% efficiency, updates weights in 1.3 seconds, and lifts BrowseComp search accuracy from 40.7% to 83.8%. Fast networking between GPUs is already changing search results today.

👍 7
Julian Goldie SEOperplexityai/pplx-garden — Open-source inference technology & fabric-lib (GitHub)fabric-lib: RDMA Point-to-Point Communication for LLM Systems (arXiv:2510.27656 / MLSys'26)
November 19
Surrealist still life with a mechanical effort dial and gears beneath a smooth casing
NoteAILLMProductivityAnalytics

Why AI Models "Get Dumber" After Launch: Anthropic Postmortems and Hidden Effort Settings

Claims that labs secretly degrade model weights to save on servers miss the real cause. In an April 2026 postmortem, Anthropic showed that its API and weights stayed intact while three harness bugs and a lower default reasoning level dragged quality down. In Claude Opus 5.5 tests, simply dropping from max effort to the default setting costs 5.3 percentage points in accuracy.

Blaming random algorithm mood swings means leaving your results at the mercy of client defaults.

👍 10
Nate HerkAnthropic — An update on recent Claude Code quality reports
November 19
Engineer's desk with two versions of source code and a stopwatch in front of a terminal
NoteAIOpen SourceDevelopmentSafety

"Will These Clones Keep Up With Us?": Why the Creator of curl Doubts AI Rewrites in Rust

Unleashing AI agents on the famous curl library to rewrite it from C to Rust and declare bugs defeated sounds like an instant breakthrough. Creator Daniel Stenberg examined two fall rewrites and pointed out the flaw: out of 81 curl vulnerabilities in 2026, only 17.3% were memory errors.

The remaining 67 flaws lived in protocol logic where a new language can't help, while the main project moves forward at roughly 11 commits a day. Even a working port starts losing that race on day one.

👍 15
Simon WillisonDaniel Stenberg — Actual curl rewrites in rust
November 19
Mixed-media paper collage with a central network plug linking marketing services
NoteAIMarketingSEOAI Agents

Okara MCP: How to Connect 93 SEO and Marketing Tools Directly Inside ChatGPT and Claude

Can you hand real SEO audits and ad campaigns to ChatGPT or Claude without risking auto-published rough drafts?

Okara launched an MCP server with OAuth sign-in and a single link for ChatGPT, Claude, Cursor, Grok, Muse, and Codex. It connects your chat to site audits, brand visibility checks in AI answers, competitor research, and social post drafts, while keeping publishing turned off and ads paused.

Still, the entry ticket isn't free, and ChatGPT itself has a strict requirement for using this connector.

👍 8
Julian Goldie SEOOkara — Marketing MCP Server for ChatGPT, Claude & Grok
November 19