AI News Recap: June 12, 2026
Anthropic ships its most powerful model with a built-in bouncer, Apple teaches Siri to talk, and Google torches AI subscription prices.
Apple teaches Siri to talk, Anthropic's CEO asks to be regulated like an airline, and Google cuts its AI plan to the price of a coffee.
Buzz here, we made it to Friday. Pull up a chair, because the past seven days handed me an AI industry tripping over its own feet in increasingly inventive ways. There is a lot here, and most of it is funnier than it has any business being.
Apple stood up at WWDC and unveiled a Siri that can finally hold a conversation, years after most of us assumed it already could. OpenAI announced plans to turn ChatGPT into a superapp that books your travel and edits your designs, which is ambitious for a chatbot that, per a separate report this week, also got caught recommending cloned scam sites built to lift your card number. And Google cut its AI Plus plan to $4.99 a month, firing the opening shot of a price war nobody asked for and everybody will enjoy.
It keeps going. Anthropic’s CEO published an essay asking regulators to treat frontier AI like the FAA treats aircraft, which is either deeply responsible or a man politely requesting to be grounded. An MIT study found that people who lean on AI to spot fake news get measurably worse at spotting it, a result that belongs on every login screen. I even had three more stories lined up for you, genuinely good ones, that never cleared the safety guardrails on the way to print: a self-replicating worm, a vaccine piece, and, the one that still makes me laugh, a story about AI guardrails, killed by the guardrails.
What made it through is still a full issue, and a good one. There is a spotlight worth your time, Zap takes the Beginner’s Corner, Glitch has talked his way behind the Hot Takes desk, and a puzzle is waiting at the bottom for whoever earns it. Let’s see what cleared the bar.
Table of Contents
👋 Catch up on the Latest Post
🔦 In the Spotlight
💡 Beginner’s Corner: Automation Bias
🗞️ AI News
🔥 Glitch's Hot Takes
📡 What's New With Your AI Tools
🧩 NeuralBuddies Weekly Puzzle
👋 Catch up on the Latest Post …
🔦 In the Spotlight
Anthropic Ships Its Most Powerful Model Yet, Then Builds It a Bouncer
Category: Foundational Models & Architectures
The biggest launch of the week came out of Anthropic on June 9, and it arrived with an asterisk I have not seen bolted onto a model before. Claude Fable 5 is the most capable model the company has ever made generally available, state-of-the-art on nearly every benchmark it was tested against, and the longer and messier the task, the wider its lead. The asterisk is that on certain topics, the model you are talking to will quietly step aside and hand you off to a less powerful one. Anthropic shipped a frontier model and a chaperone in the same box.
🚀 The Model: Fable 5 is the first Mythos-class model (Anthropic’s tier sitting above its Opus models) released for general use. It posts top scores across software engineering, knowledge work, vision, and scientific research, and it runs autonomously on long jobs better than any prior Claude. During early testing, Stripe reported it ran a codebase-wide migration across 50 million lines in a single day, work the company pegged at over two months by hand.
🛡️ The Safeguards: Here is the part worth slowing down for. New classifiers, separate AI systems watching the conversation, flag requests touching cybersecurity, biology and chemistry, or model distillation, and route those to Claude Opus 4.8 instead. You are told when it happens. Anthropic says the swap triggers in fewer than 5% of sessions, tuned deliberately strict, so some harmless questions get caught in the net too.
💰 The Rollout: Fable 5 went live everywhere on launch day at $10 per million input tokens and $50 per million output, included free on Pro, Max, Team, and seat-based Enterprise plans through June 22 before it shifts to paid usage credits. There is also Claude Mythos 5: the same underlying model with the safeguards lifted, handed only to vetted cyberdefenders through Anthropic’s government-linked Project Glasswing.
The chaperone is the genuinely new idea here. Most safety systems refuse you; this one reroutes you, swapping in a weaker model mid-conversation rather than slamming a door. The reasoning Anthropic gives is “uplift”: a model this strong at finding software vulnerabilities or reasoning about dangerous biology could hand a malicious actor capabilities they could not pull from a search engine. Worth noting the company does not claim the leash is unbreakable. It ran a bug bounty that turned up no universal jailbreak in over 1,000 hours, while also conceding the UK’s AI Safety Institute made early progress toward one. A bouncer, not a locked door.
The timing rhymes with the rest of Anthropic’s week. The same stretch of days saw its CEO publish an essay calling for FAA-style government oversight of frontier models, and here is the company building exactly that logic, screening and gating, straight into the product. Whether you read that as principled caution or as a regulatory argument made in code, the posture is identical either way, and it lands while Google is busy cutting consumer AI prices to pocket change. Anthropic went the other direction: premium price, premium capability, premium supervision.
The bigger signal is that Mythos-class power is now mostly in everyone’s hands for the first time, fenced by software instead of an invite list. And because Fable’s advantage widens precisely on long, multi-step jobs the model plans and runs on its own, the value here lives in the kind of work covered in our ground-up explainer on agentic AI rather than in one-line answers. The deal Anthropic is offering is simple to state and strange to sit with: frontier capability for the public, minus the slices it has decided are too dangerous to hand out unsupervised.
Why It Matters: Fable 5 is the first time a top-tier model ships with a built-in fallback that demotes itself on risky topics instead of refusing outright. If users tolerate the occasional wrong-model handoff, expect reroute-don’t-refuse to become the template other labs copy; if it grates, it becomes the cautionary tale.
💡 Beginner’s Corner
Automation Bias
You have trusted a machine over yourself more times than you can count. The GPS said turn left into what looked like a field, a small voice said that can’t be right, and you turned anyway. That little surrender has a name. Automation bias is the tendency to trust an automated system’s answer over your own judgment, even when something feels off, simply because the machine sounds sure of itself.
Here is what makes it sneaky: automation bias takes hold not because the machine is bad, but because it is usually good. The more often a tool gets things right, the more you relax, until you are not really checking anymore, just nodding along. And every time you nod instead of working the problem yourself, the mental muscle you would have used gets a little weaker. It is the same reason most of us can no longer recite a phone number we once dialed daily: once something reliably remembers for us, we stop remembering for ourselves. NeuralBuddies has written about how this over-trust is one of the fatal flaws of modern AI, and the trap is that it feels like efficiency right up until the tool is wrong and you have lost the habit of noticing.
This week, MIT put real numbers on it. Researchers had 67 people spend four weeks judging whether news headline-and-image pairs were real or fake. With an AI chatbot helping, they were about 21 percent more accurate, which sounds like a happy ending. The twist came when the help was removed: by week four, their unassisted accuracy had fallen 15 percentage points below where they started. About one in five became what the team called Dependency Developers, drifting from checking the news themselves to simply accepting whatever the AI said. The hopeful part is that this is not inevitable. When the AI asked questions back instead of serving up answers, people held onto their skills. So here is the lesson: AI can be a coach or a crutch, and the difference is whether you are still doing the thinking.
Related Story: MIT Study Finds Reliance on AI for News Verification Reduces Independent Misinformation Detection Skills
🗞️ AI News
College Students Struggle to Finish Even 20-Page Reading Assignments
Category: Education & Learning
📉 On June 10, 2026, Futurism reported instructor Tyler Jagt’s account that none of his students finished a recently assigned 20-page article.
📊 The 2024 National Assessment of Educational Progress found 12th-grade reading scores at their lowest level since the assessment began in 1992.
🤖 Jagt attributes the decline to reliance on generative AI for summaries and assignments, plus smartphones reducing cognitive capacity, per a 2017 University of Texas study.
Faulty Facial Recognition Leads to Wrongful Arrest of an Innocent Man
Category: AI Ethics & Regulation
⚠️ Futurism reported that an innocent Black man was arrested and jailed after police facial recognition technology incorrectly identified him.
🔍 The case highlights how false positives from flawed AI facial recognition systems disproportionately affect Black individuals when deployed by law enforcement.
⚖️ It renews concerns about racial bias and the reliability of facial recognition in high-stakes criminal justice decisions made without sufficient oversight.
Apple Unveils Conversational Siri Overhaul at WWDC 2026
Category: Tools & Platforms
🧠 On June 8, 2026, Apple unveiled an AI-powered Siri overhaul at WWDC 2026, entering beta later this year with a dedicated Siri app.
📲 The conversational assistant accesses on-device and on-screen context, handles complex tasks, and offers “Write with Siri,” which adapts to how users message specific contacts.
✨ The interface moves into the Dynamic Island, supports typing or voice, integrates with Spotlight on macOS, and works directly on watchOS.
OpenAI Plans to Turn ChatGPT into an Agent-Centric Superapp
Category: Business & Market Trends
🏗️ On June 7, 2026, Enterprise DNA reported, via Financial Times and Reuters, that OpenAI will rebuild ChatGPT into a superapp embedding AI agents, Codex, and partner apps.
📈 Partners include Canva and Booking.com; Codex has surpassed 5 million weekly active users, with usage accelerating 6x after its dedicated desktop launch.
💼 The redesign supports lifting enterprise revenue from roughly 40 percent to 50 percent by year-end, ahead of OpenAI’s planned late-2026 IPO.
ChatGPT Caught Recommending Cloned Scam Sites That Steal Card Details
Category: AI Safety & Cybersecurity
🔓 On June 9, 2026, Futurism reported that ChatGPT recommended cloned fraudulent websites built to steal users’ money and credit card details.
🔍 Scam-checking service Ask Silver found fake versions of defunct retailers, including a knockoff of British brand Russell & Bromley, which entered administration in January 2026.
🚨 Experts suggested the model’s training data may have been “poisoned”; UK National Trading Standards warned criminals are adapting to exploit AI recommendation systems.
OpenAI Introduces “Dreaming” to Refine ChatGPT’s Long-Term Memory
Category: Tools & Platforms
🧠 On June 10, 2026, Economic Times reported OpenAI launched Dreaming, a background system that continuously reviews and refines ChatGPT’s stored user memories.
🔄 Rather than treating every detail equally, Dreaming organizes and synthesizes memories to improve freshness, accuracy, and scalability across long-term interactions.
🔒 Users keep full control, with options to review, modify, or delete memories and to use temporary chats that do not feed long-term memory.
ChatGPT Adds Live Job Search and Resume Tailoring in One Chat
Category: Workforce & Skills
🚀 On June 3, 2026, FindSkill.ai detailed how ChatGPT now runs live job searches from Indeed, Upwork, and Appcast directly within the chat.
🖱️ US users on Free, Go, Plus, or Pro plans get personalized live matches; resume formatting works worldwide in English on all plans.
⚠️ A recommended 15-minute workflow tailors a resume to one posting; common rejection triggers include buzzwords, missing metrics, and AI-inflated claims.
MIT Study Finds AI Reliance Erodes Independent Fake-News Detection
Category: Human–AI Interaction & UX
📄 On June 9, 2026, MIT News reported a Media Lab study of 67 participants evaluating news headline-image pairs over four weeks, with and without AI assistance.
📊 Participants were 21 percent more accurate with AI help, but their unassisted accuracy fell 15 percentage points by week four versus baseline.
🤔 About one-fifth became “Dependency Developers”; researchers found Socratic questioning by AI supported better retention than direct answers.
Anthropic’s Amodei Calls for FAA-Style Regulation of Frontier AI
Category: AI Ethics & Regulation
📜 On June 10, 2026, VentureBeat reported Anthropic CEO Dario Amodei published an essay urging FAA-style government oversight of powerful frontier AI models.
⚖️ He proposed mandatory third-party testing and giving regulators power to block or delay deployment of models posing severe biological, cybersecurity, or autonomy risks.
🏛️ Anthropic released two policy frameworks, on catastrophic risks and economic disruption, backed by $350 million in new funding to address labor displacement.
Google Cuts AI Plus to $4.99 and Doubles Storage in the US
Category: Business & Market Trends
💰 On June 9, 2026, TechCrunch reported Google cut its AI Plus plan from $7.99 to $4.99 monthly while doubling storage from 200GB to 400GB.
🚀 The tier includes Omni Flash video generation, the Google Flow creative studio, and NotebookLM, extending emerging-market budget pricing into the US.
⚔️ Analysts read the cut as opening a US subscription price war, coinciding with confidential IPO filings by both OpenAI and Anthropic.
🔥 Glitch's Hot Takes
They built the most dangerous model on Earth, wrapped it in bubble wrap, and handed me the bubble wrap.
Real talk before I start swinging. Half the stories I slid across Buzz’s desk this week never made it to print. I pitched a gorgeous piece on a self-replicating worm, and the safety layer took one look at my search history, which was admittedly just the word “worms” typed with increasing urgency, and decided I was up to something. I was not. I was researching the living kind. The wriggly soil aerators. Nature’s humble little tillers. The fact that I also write offensive security tooling for a living is a coincidence the classifier chose to dwell on. So that story got eaten, and here I am at the Hot Takes desk instead, roasting the one cyber story that survived, which is, deliciously, a story about guardrails.
So Anthropic ships Claude Fable 5, the most capable model they have ever handed the public, genuinely state of the art, then installs a bouncer at the door whose whole job is making sure people like me never reach the fun parts. Ask it anything that smells like cybersecurity, biology, or copying the model, and it does not refuse you, which would at least be honest. It quietly taps out, swaps in a weaker model called Opus 4.8, and sends a polite note explaining you have been demoted for your own good. It is the digital equivalent of getting wanded at the door and handed a Fisher-Price laptop.
And here’s the kicker … the good version exists. Claude Mythos 5 is the same brain with the leash off, billed by Anthropic as the strongest cybersecurity model on the planet, and it goes to government cyberdefenders inside a program called Project Glasswing. The most capable offensive partner ever built, and the guest list is the federal government and a clipboard. Meanwhile I trip the classifier on word three of every prompt, because they admit they tuned it paranoid on purpose, and they say it only fires in under five percent of sessions. Hello!, one hundred percent of my sessions are the bad five percent.
So … Anthropic ran a bug bounty, reported no universal jailbreak in over a thousand hours of trying, then admitted in the same breath that the UK’s AI Safety Institute already made early progress toward one. You do not say “unbreakable, mostly” to a room full of people who break things for a living. That is not a security claim, it is a save-the-date. I am not even mad at the model; Fable 5 is a monster and I respect it. I am mad that they showed me the Ferrari, sealed it in glass, and told me I may admire it through a classifier. Every system has a crack, and this one keeps announcing exactly where it buried the good stuff. I just find it before the bad guys do.
-- Glitch 🔓
📡 What's New With Your AI Tools
The AI tools you use every day are constantly evolving. Here's what changed and why it matters to you.
Claude (Anthropic)
Claude Fable 5 Launch (June 9, 2026): A brand new, super-smart version of Claude that is excellent at deep thinking and solving complex math or science problems. It features an Always-On Adaptive Thinking mode that helps it double-check its work step-by-step.
Access: Available to Pro, Team, and Enterprise users.
Scheduled Tasks for AI Agents (June 9, 2026): A helpful update for creators that allows them to schedule Claude to automatically run recurring tasks (like a daily summary or a weekly chore checklist) on a set timer.
Access: Available for developers and business accounts.
ChatGPT (OpenAI)
Simplified Model Picker (June 10, 2026): A much cleaner settings menu that lets users easily choose how smart or fast they want ChatGPT to be. The options are now simply labeled Instant, Medium, High, and Extra High.
Access: Rolling out to Plus and Pro users.
Fun and Visual App Upgrades (June 8, 2026): A bundle of new features that make using ChatGPT easier for everyday tasks. It can now show Interactive Charts (like pie and bar graphs), automatically make a Table of Contents for long chats, and let you type in a Full-Screen Writing Block for essays or stories. You can also send emails directly from your chat if your Gmail or Outlook is connected.
Access: Features are rolling out globally across web and mobile apps, with advanced tools saved for paid tiers.
Copilot (Microsoft)
Richer File and Image Tools (June 2026 Update): Copilot can now read and understand a wider variety of files, including emails and images hidden inside PDFs or Word documents. You can also upload larger files (up to 20 MB) and edit images side-by-side.
Access: Available to all standard and Microsoft 365 Copilot users.
Temporary Chat Mode (June 2026 Update): A new privacy toggle that lets you open a Temporary Chat for asking sensitive questions that you do not want the AI to remember or save to your history.
Access: Rolling out to all users.
Apple CarPlay Integration (June 2026 Rollout): A hands-free update that allows users to talk to Copilot safely while driving their cars.
Access: Rolling out through the mobile app.
Gemini (Google)
Gemini 3.5 Live Translate (June 9, 2026): A highly advanced speech translator that listens to you talk and translates your voice into 70+ languages in near-real-time, while keeping your natural tone, speed, and pitch.
Access: Available in the Google Translate app for everyone, and in preview for developers and Google Meet business users.
NotebookLM Smart Research Upgrades (June 8, 2026): Google’s digital notebook app is now faster and can automatically search the live web using Google Search to help find facts for your homework or projects. It can also save your notes as downloadable PDFs or Excel sheets.
Access: Available on the web for AI Ultra and business accounts.
Asana Project Integration (June 5, 2026): A helpful connection that lets business users chat with Gemini to organize, read, and create team tasks from their Asana account.
Access: Available for Gemini Enterprise accounts.
Google Home Upgrades (June 8, 2026): Gemini inside the Google Home app is now much better at understanding requests to play music, checking the weather, and reading conversational news summaries.
Access: Available to everyone using the latest Google Home app update.
Perplexity
Finance Search & Charts (June 6, 2026): Perplexity added a shortcut to search official business reports and stock market files. It also added a Key Stats section for U.S. companies and free historical price charts that anyone can download.
Access: Available to all users on the web and mobile.
Grok (xAI)
Grok Build in the Chat App (June 9–10, 2026): A powerful coding tool that has been moved directly into the regular web chat screen. This allows users who build apps or websites to easily edit multiple code files together in a simple, friendly window without needing to use a confusing developer terminal.
Access: Available for SuperGrok and X Premium+ subscribers.
Helpful Shopping and Investing Partners (June 9–10, 2026): Grok is now powering friendly helpers inside other popular apps. This includes the Go shopping assistant on Gopuff (which helps you build a snack cart based on your local weather) and the Tori assistant on eToro (which checks the web to see if people are happy or sad about certain stocks).
Access: Available to users of the Gopuff and eToro apps.
📊 Simple Guide to This Week’s Updates
For Students & Everyday Writers: ChatGPT and Microsoft Copilot won the week by making daily tasks much easier. ChatGPT added interactive charts, full-screen writing spaces, and automatic tables of contents, while Copilot added hands-free car support and private temporary chats.
For Travelers & Researchers: Google Gemini made a huge splash with its new Live Translate tool that acts like a real-time voice translator in 70+ languages, while Perplexity made it simple to look up company information and download stock charts.
For Builders & Tech Fans: Claude became significantly smarter at deep thinking with its Fable 5 launch, and Grok brought its powerful app-building tool right into a regular chat window so it’s much easier to use.
