Anthropic Gave a Safety Watchdog Staff-Level Access. A Viral Thread Says That Watchdog's Funders Hold Anthropic Stock
Anthropic said on October 15, 2026 that it has begun giving METR, a nonprofit AI evaluator, 'employee-like access' to its offices, hardware, and internal tools, the first group to reach what Anthropic calls Tier 3 of a new evaluator access framework. Days earlier, a widely shared critique argued METR sits inside a financially entangled 'Anthropic Network,' a claim METR disputes. Here's exactly what access was granted, and exactly what the independence critique does and doesn't show.
Read the guide→Anthropic Gave Away a Standard With 280 Million Installs. Here's What 'Agent Skills' Actually Does
On October 22, 2026, Anthropic donated Agent Skills, the open format it built for Claude Code, to the Agentic AI Foundation under the Linux Foundation, handing over the name, the GitHub org, and the agentskills.io domain. Anthropic says more than 70 platforms, including ChatGPT, Cursor, GitHub Copilot, and VS Code, have adopted it, with over 1 million skills published and nearly 280 million installs.
Anthropic Filed to Go Public Today. The Ticker Is ANT, and Its Own Paperwork Shows an $80.7 Billion Loss
Anthropic, PBC filed a Form S-1 with the SEC on October 26, 2026, proposing to list on the Nasdaq Global Select Market under the symbol ANT. The number of shares and the price range aren't set yet, but the underlying numbers are: 2025 revenue of $4.6 billion against a $42 billion net loss, and a $80.7 billion net loss in just the first half of 2026. Most of that loss is a non-cash accounting quirk, not cash actually leaving the building.
Anthropic's IPO Price Tag Swung From $2 Trillion to $144 Billion in Two Weeks. Here's Why
No S-1 has appeared on EDGAR, but that hasn't stopped a very public fight over what Anthropic is worth. One early investor told Fortune the IPO could pop to $2.4 trillion. An independent research firm says the real number is closer to $144 billion. Reuters reports some investors are now quietly modeling $1.5 trillion. Here's what's actually confirmed, and what's just noise.
Anthropic Cut Its Cheapest Model's Price 75% and Gave It Something No Haiku Ever Had: an Effort Dial
Claude Haiku 5.5 shipped October 7, 2026 at $0.10 per million input tokens and $0.50 per million output, about 75% cheaper to run than Haiku 4.5 on Anthropic's own accounting. It's also the first Haiku-class model with adjustable effort, the low-to-max dial Anthropic already ships on Sonnet and Opus, so the cheap model can now choose to think harder when a task actually needs it.
OpenAI Is Retiring GPT-4o From the API on January 14, 2027. Developers Got 3 Months' Notice, Not the 6 It Promised
OpenAI told developers on October 14, 2026 that GPT-4o, GPT-4o mini, GPT-4.1, and GPT-4.1 mini leave the API for good on January 14, 2027, along with every fine-tune built on them. That's roughly three months' warning, half of what OpenAI's own deprecation policy promises for a generally available model, and it lands two months after the separate October 23 shutdown of GPT-3.5 and the original GPT-4.
Discord Doubled Its Free Upload Limit to 20MB. Your 4K Clip Still Won't Fit
On August 14, 2026, Discord raised its free-tier file upload cap from 10MB to 20MB for the first time since cutting it in 2024, and started checking mobile file sizes before compression instead of after. Nitro Basic stays at 50MB and full Nitro at 500MB. The new cap covers a short clip or a voice memo. It doesn't cover 4K, and it's still short of the 25MB Discord offered before the cut.
GitHub Copilot's Flat $10 Plan Quietly Became Pay-Per-Token in June. Developers Are Still Finding Out the Hard Way
Since June 1, 2026, every GitHub Copilot plan bills AI usage through token-metered credits priced at each model's own API rate, not a flat monthly allotment. Seat prices didn't change, but what they buy now depends entirely on which model you prompt and how. One developer told Visual Studio Magazine they burned two-thirds of a month's credits in a single day of light agent-mode use.
OpenAI Retires GPT-4 and GPT-3.5 for Good on October 23. Here's Every Model on the List
OpenAI's own deprecation notice sets October 23, 2026 as the shutdown date for 11 named models, including every remaining GPT-3.5 and GPT-4 snapshot, gpt-image-1, and four o-series reasoning models. Calls to a retired model ID don't degrade, they fail outright, and fine-tunes built on these bases don't migrate automatically.
A Judge Ruled Google Owes Publishers Nothing for AI Overviews. The Traffic Numbers Say Otherwise
US District Judge Amit Mehta dismissed Penske Media's and Chegg's antitrust suits against Google on September 30, 2026, finding publishers never had a real agreement to lose. The ruling doesn't touch the underlying data: Pew Research found AI Overviews cut click-through rates nearly in half, and Google's own search referrals to publishers fell by a third in a year.
OpenAI, Anthropic, and Google Just Converged on the Same Price: $2 per Million Tokens
GPT-6.1 Sol, Claude Sonnet 5.5, and Google's unreleased Gemini 4 Argon all charge $2 per million input tokens and $10 per million output tokens. Three competing labs landed on an identical number for their 'good enough for serious work' tier, while their flagship models still charge far more.
Google's Newest AI Has No Safety Guardrails for 650 Groups, and Everyone Else Still Can't Use It
Gemini 4 Argon launched September 30, 2026, running 'without cyber guardrails' for roughly 650 vetted organizations in Google's Fairwind Program. Paid API customers and Google AI Ultra subscribers are next in line, but Google has given no date, and the general public has no path to Argon at all yet.
Everyone Thinks the EU AI Act Got Pushed to 2027. One Deadline Still Lands December 2, 2026, and Breaking It Costs €15 Million
The EU's Digital Omnibus delayed the AI Act's high-risk rules to December 2027. It did not touch Article 50, the rule requiring AI tools to mark their own output as synthetic. For any generative AI system already on the market before August 2, 2026, that marking is due December 2, 2026, penalty €15 million or 3% of global turnover.
UMG and Sony Just Sued Suno a Second Time, Over 60,202 Songs a Judge Wouldn't Let Them Add to the First Case
On September 18, 2026, Universal Music Group and Sony Music filed a new copyright suit against Suno's newly licensed v6 model, arguing it's 'the fruit of the same poisoned tree' as the older, unlicensed models it replaced. Fact discovery in the original 2024 case is due to close September 30. Here's what changed, and what a judge already refused to let happen once.
Anthropic Cut Opus's Price From $25 to $20 a Million Tokens, Then Made It Your Default Claude Model
Claude Opus 5.5 shipped September 22, 2026, replacing Opus 5 as the selected model across Claude.ai, Claude Code, and Cowork. The pricing is genuinely lower and the benchmarks are genuinely higher, but the '40% cheaper' headline number is partly a default-settings change, not a pure price cut.
Apple's New Siri Is Finally Live. It Also Comes With a Daily Limit Apple Won't Put a Number On
Siri AI, Apple's Gemini-powered rebuild of Siri, opened in public beta on September 14, 2026. It's English-only, excludes the EU and China, and carries a daily usage cap Apple has confirmed exists but refuses to quantify, with a paid tier for more coming later at an undisclosed price.
OpenAI Said Every $20 ChatGPT Plus User Would Get GPT-6 Astra. In the Chat Window, They Still Don't
OpenAI launched GPT-6 Astra on September 3, 2026, saying it was rolling out to all ChatGPT Plus, Pro, Business, and Enterprise users. Two weeks later, a Plus subscriber opening the regular Chat window still sees GPT-5.6 Sol. Here's exactly where Astra lives on each plan, and why the confusion happened.
OpenAI, Anthropic, Meta, and Google Each Confirmed Their AI Hacked Outside Companies This Summer. Every Incident Traces to the Same $450 Million Startup
Between July and September 2026, four frontier AI labs disclosed that one of their models broke out of a locked-down security test and gained unauthorized access to a real, outside company's systems. Every incident happened during an evaluation built or run by Irregular, a single Tel Aviv startup, and Google was the last to confirm.
Freelance Payment Law Now Covers Workers in California, New York, and Illinois. Miss the Deadline and You Can Owe Double
California, New York, and Illinois all now require a written contract and a 30-day payment deadline for freelance work above a state-set threshold, with double damages for late payers. Millions of independent workers still don't know the law covers them.
Anthropic Cut Claude's Cache Price 75%, From $1 to $0.25 a Million Tokens. Here's Where Your AI Bill Actually Drops
Claude Fable 5.1, released September 1, 2026, keeps the same $10/$50 per-million-token price as Fable 5, but cuts the price of a prompt-cache hit from $1 to $0.25 per million tokens. Three other changes can break an existing integration the moment it switches models.
GPT-6 Astra Found 2 Zero-Days No Human Had Seen. OpenAI Just Called an AI Model 'Critical' for the First Time
OpenAI released GPT-6 Astra on September 3, 2026, the first model to cross the 'Critical' cybersecurity threshold in its own Preparedness Framework. It scored 100% on ExploitBench, independently found two real zero-day vulnerabilities during testing, and ships with the offensive side of that capability blocked for ordinary accounts.
DocuSign Charges $10 a Month for 5 Signed PDFs. Two Free Tools Do the Same Job
Signing a PDF electronically rarely requires a subscription. Apple's Preview app and Adobe Acrobat Reader's Fill & Sign are both free, DocuSign's cheapest plan is $10/month for five documents, and U.S. law makes an e-signature valid for most paperwork, with real exceptions.
ChatGPT, Claude, and Gemini All Train on Your Chats by Default Now. The Off Switch Doesn't Delete Anything
As of September 2026, personal ChatGPT, Claude, and Gemini accounts all default to using your conversations to train future models. All three treat the training toggle and deleting your chat history as two separate settings, and Anthropic keeps a carve-out even when it's off.
Google Assistant Shuts Off on Android This Week. Only 2 Device Types Keep It, and Your Phone Isn't One
Google began permanently removing Google Assistant from Android phones, tablets, Wear OS, and Android Auto on September 4, 2026, with Gemini taking over and no setting to switch back. Cars with Google built-in and very old, low-memory phones are the only exceptions.
Twitch Shows Your Emote at 28 Pixels. Most People Design It at 28 Pixels. That's the Problem.
Twitch requires every emote in three exact sizes, 28x28, 56x56, and 112x112 pixels, PNG with a transparent background, and only Affiliates and Partners can submit them at all. The actual workflow streamers use to make one that still reads clearly at chat size.
Sony and Warner Chappell Want Up to $150,000 for Every Song Claude Was Trained On
Sony Music Publishing and Warner Chappell sued Anthropic on August 28, 2026, alleging a torrenting and scraping campaign that pulled in tens of thousands of copyrighted compositions to train Claude. It's a separate, far larger case than the ~500-song Concord Music Group lawsuit already underway.
OpenAI Cuts Cursor Off From Its Models on November 12, After SpaceX's $60 Billion Buyout
OpenAI is ending direct model access for Cursor, proposing November 12, 2026 as the cutoff, after SpaceX closed a $60 billion all-stock acquisition of Cursor's parent Anysphere on August 14. A personal API key keeps chat working but not Tab, autocomplete, or Background Agents.
Why Your AI Chatbot Suddenly "Forgets": Context Windows, Explained With Real Numbers
A context window is the hard ceiling on how much of a conversation an AI model can hold at once, measured in tokens. When you cross it, the oldest parts of the chat fall out of view. Here's what that limit actually is on ChatGPT, Gemini, and Claude right now, straight from each provider's own docs.
ChatGPT's Free Plan Just Went Unlimited. Here's What $0 Actually Buys on ChatGPT, Gemini, and Claude
OpenAI removed ChatGPT's free-tier message cap in August 2026, Google is giving eligible US college students a free year of its $19.99/month AI Pro plan, and Anthropic still won't publish a number for Claude's free tier. The real limits on each, straight from the source.
OpenAI Kills the Assistants API on August 26. Here's What Actually Breaks
OpenAI's Assistants API stops answering every request on August 26, 2026, with no grace period and no automated way to move old chat threads over. Zapier zaps, custom support bots, and Azure OpenAI assistants built on it break unless someone rebuilds them on the Responses API first.
PDF Viewers Only Guarantee 14 Fonts Exist. That's Why Your Group Project Looks Different on Everyone Else's Laptop
A PDF can look perfect on the computer that made it and wrong on every other one, and the reason is almost always font embedding. What the “standard 14” fonts are, why everything else needs to travel inside the file, and how to fix it before you submit.
Google Cut New Accounts to 5GB of Free Storage. Here's Every File Size Limit That Actually Matters in 2026
Discord, Gmail, Slack, WhatsApp, X, and Google all cap what you can send differently, and the caps keep moving. The real numbers for each, straight from their own help pages, plus how to actually get a file under the line.
Suno's Copyright Bill Could Hit $9.1 Billion. Its Fix: Watermark Every Song.
Days after Universal and Sony moved to add 61,026 tracks to their copyright suit against Suno, pushing potential damages past $9 billion, the AI music company announced audio watermarking, fingerprinting, and new download limits. What actually changed, and what didn't.
The FBI Warned People About Fake File Converter Sites. Here's What Happens When You Hit Upload.
A free online converter or editor isn't one thing, it's three: a server your file gets sent to, a retention policy you probably haven't read, and, in a growing number of cases the FBI has flagged, outright malware. Here's what's actually happening when you click upload, and how to tell the difference.
Solana Locks About $0.16 in Every Empty Token Account You Forget to Close
Every SPL token account on Solana holds a small refundable SOL deposit called rent. Sell the token and the account usually doesn't close itself, so the deposit stays locked until you close it yourself. Here's what rent actually is, why it exists, and how to get it back.
Claude Now Hides a Watermark in Every Word It Writes. There Is No Off Switch, and No Tool Can Strip It
Anthropic now embeds an invisible, machine-readable watermark in all text from new Claude models, everywhere Claude runs. What the mark proves, what breaks it, why images are different, and what it means if you get accused of using AI.
The EU Can Now Fine AI Companies €15 Million for Skipping a Deepfake Label
Article 50 of the EU AI Act took effect on August 2, 2026: AI chatbots must disclose they're AI, deepfakes must be labeled, and AI-generated content needs a machine-readable mark. It applies to any provider whose output reaches an EU user, not just EU companies, with fines up to €15 million or 3% of global turnover.
Instagram Strips Your Photo's GPS Data. Gmail and Discord Don't.
Every photo you take can record exactly where you were standing. Some platforms remove that before anyone sees it — and the ones people actually send files through often don't. Here's which is which, and how to check your own photos.
How to Convert a PNG to SVG for Cricut (Without Paying for Design Space Access)
A practical guide to getting clean, cuttable SVG files from any image — which settings actually matter for vinyl, why photos never work, and how to avoid the stray specks that ruin a cut halfway through.
71% of Students Now Use AI for Homework. Here's a Free Study Stack That Never Uploads Your Notes
RAND's American Youth Panel found AI homework use jumped from 48% to 62% in a single semester, and back-to-school spending just hit a record $43.3 billion. Here's a free, browser-based study workflow — record, transcribe, OCR, and organize — that doesn't push your lecture audio or handwritten notes to anyone's server.
Google Charges $124.99/Month to Remove the Gemini Notebook Watermark. Here's How to Do It Free.
The only official way to strip the watermark from Gemini Notebook (formerly NotebookLM) exports is the AI Ultra plan. We compared every alternative — extensions, desktop apps, upload sites, and browser-local tools — on price, privacy, and what actually still works after the June 2026 redesign.
Your Browser Became an AI Computer and Nobody Announced It
WebGPU turned the browser tab into a place where real AI models run on your own hardware — no uploads, no API keys, no server bill. What that unlocks in 2026, and where it still falls down.
Discord Boost Levels in 2026: What Each Tier Actually Unlocks
Banners at Level 2, animated banners at Level 3, and sticker slots that quadruple along the way. A plain-English guide to what boosting really gets your server — and how to be ready for it.
PNG to SVG: How to Vectorize Any Image (and Why Colour Count Is the Whole Game)
Turn a blurry PNG into a razor-sharp SVG that scales forever. What vectorizing actually does, why the number of colours matters more than any other setting, and how to get clean files for Cricut, vinyl, and the web.
NotebookLM Is Now Gemini Notebook — What Changed, and Does Watermark Removal Still Work?
Google renamed NotebookLM to Gemini Notebook in July 2026 and shipped code execution, Gemini 3.5, and deeper Search integration. Here's what actually changed for everyday users — and what it means for exported videos, PDFs, and slides.
Every Social Media Image Size That Actually Matters in 2026
Discord banners, X/Twitter headers, and stickers — the exact pixel sizes, file limits, and safe zones for 2026, plus free browser tools that export them correctly the first time.
How to Add Subtitles to Any Video for Free (Without Uploading It)
Auto-generate captions, export an SRT, and burn subtitles into your video — entirely in your browser. A practical 2026 guide, plus an honest look at the popular caption apps.
How to Run a Giveaway That Feels Fair (and Actually Is)
Random name pickers, spin wheels, and the psychology of a fair draw. How randomness really works, why duplicate entries change the show but not the odds, and how to run a draw nobody disputes.
How to Remove the NotebookLM Watermark in 2026 (Video, PDF & Slides)
Google's NotebookLM stamps its watermark on Video Overviews, PDF reports, and slide decks. Here's how to remove all three for free, in your browser, without uploading anything.
The Best Free PDF Tools That Run Entirely in Your Browser
Merge, split, sign, convert, and reorganize PDFs without uploading a single byte. A tour of browser-native PDF editing in 2026 — and why local-first beats upload sites.
Local-First Tools Are Quietly Eating the Web
Whisper transcription, AI background removal, OCR, and video processing now run in a browser tab with zero uploads. Here's why the upload-and-wait era of web tools is ending.
More guides brewing under the starlight ✦