Welcome, humans.
Webcomic creator Sean Kleefeld logged into Amazon on Monday and found 15 years of order history, his Comixology library, his Prime membership, and every dollar of income tied to his self-published books… gone. No flag. No appeal. No human on the other end.
What happened: Amazon deployed an AI agent that, rather than flagging accounts for potential violations, just cancels them. Outright. Another creator, Tom Ray, reportedly lost his entire "per-page" formatted comics catalog going back to 2018. When an 800-pound gorilla outsources due process to a moderation LLM, losing one platform can technically mean losing a life's work.
Everyone worried about AI taking your job? Guess what. AI just figured out how to take your account. If Claude Mythos gets loose, hackers could take your bank account next.
Here’s what happened in AI today:
😻 Anthropic shipped Claude Opus 4.7; the vision leap is real, the tokenizer change isn't free.
📰 OpenAI overhauled Codex into a full Mac-level agent workstation with 90+ plugins.
📰 Anthropic's CPO quit Figma's board as Anthropic preps competing design software.
🍪 Canva rebranded as "an AI platform with design tools" at its $42B IPO test.
🎓 Nest CLAUDE.md files at every folder level so Claude Code auto-loads your context.
P.S: Want to reach 700,000 AI-hungry readers? Click here to advertise with us.
P.P.S: Love robots? We’re starting a new robotics newsletter! Sign up early here.
😻 Opus 4.7 Dropped. OpenAI Overhauled Codex. On The Same Day. Classic ThursdAI, am I right?
Yesterday's AI calendar bent around two flagship launches within hours of each other. If you do any agentic coding work, your stack just got upgraded. But fair warning: the time it takes you to hit your rate limits probably shrank too too.
Here's what happened:
Anthropic shipped Claude Opus 4.7 at the same $5/$25 per million token pricing as 4.6.
Visual reasoning (how well AI sees) jumped from 69.1% to 82.1%, with images processing at up to 2,576 pixels on the long edge (over 3x any prior Claude).
SWE-bench Pro (coding benchmark) went from 53.4% to 64.3%, and Opus 4.7 is now #1 on Vals AI's Vibe Code Benchmark at 71%.
The catch buried in Anthropic's own docs: the new tokenizer can use up to 35% more tokens for the same text. Combined with Claude Code's new xhigh effort default, Pro and Max users will hit weekly caps faster unless they manually dial down.
Hours after Anthropic’s release… OpenAI overhauled Codex, its coding app, to be “Codex for (almost) everything”… a.k.a OpenAi’s version of Claude Cowork. It comes with Mac-level computer use (agents click and type alongside you), an in-app browser, persistent memory, automations that wake up across days, and 90+ new plugins (Atlassian Rovo, CircleCI, Microsoft Suite, etc). It can do a lot more than that too.
Keeping score on the SaaSpocalypse: Anthropic CPO Mike Krieger resigned from Figma's board the same day reports surfaced Anthropic is shipping design software. Figma stock slid. Rumor has it that was going to launch yesterday… but it did not.
Pro tips (straight from the Claude Code team): Boris Cherny and Cat Wu both shipped launch-day playbooks on how to actually use 4.7. The unifying framing, from Wu: treat it like an engineer you're delegating to, not a pair programmer you're guiding line by line. The four moves that matter:
Front-load your context. Goal, constraints, & acceptance criteria (more below), all in the first turn. 4.7 is built to take a full brief and run with it; give it a vague goal and you'll get 4.6-level results from a model capable of more.
Turn on auto mode. Hit Shift+Tab in the terminal (Max, Team, Enterprise). Your permission prompts will route through a safety classifier, so you can run parallel Claudes to get more work done without babysitting any single run.
Tell it how to verify its own work. Both Cherny and Wu call this the 2-3× output-quality multiplier. Put your testing workflow in your
claude.mdso 4.7 runs tests every time, or install a/verify-appskill for your stack.xhighis the new Claude Code default. Use/effortto step down on routine work; savemaxfor the hardest tasks (it's session-only, so it doesn't persist into your next session).
In the web app, you now have this thing called adaptive thinking that lets Claude decide how long to think for. In other words: it’s a thinking router? Booo! Well, it’ll take time to “adapt” to this, but for now, use it because otherwise it won’t think much at all.
Our take: The story from what we’ve read is the tokenizer. Anthropic's "same pricing as 4.6" is technically true; in practice, the same prompt maps to more tokens, the default effort runs higher, and the output is longer. A friend of ours hit their weekly Max limit in basically one prompt.
That said, Opus 4.7 is a better model. It's also a more expensive one, even with unchanged sticker prices. The people who get the most out of it won't be the ones running at defaults. They'll be the ones who actually follow the best practices. Study up.
FROM OUR PARTNERS
Most tools are still locked to their own database, blind to everything users already have in Slack, GitHub, Salesforce, Google Drive, and dozens of other apps. That's the ceiling on what you can build.
WorkOS Pipes removes it. One API call connects your product to the apps your users live in. Pull context from their tools, cross-reference data across silos, power AI agents that act across services. All with fresh, managed credentials you never have to think about.
🎓 AI Skill of the Day: Nest a CLAUDE.md at Every Level of Your File Structure
Taylor Pearson shared the best Claude Code tip we've seen all month: Claude Code automatically loads the CLAUDE.md file from every directory above whatever file you're working on. Nest those files intelligently and every new chat starts with full context already loaded.
His structure:
Global (~/CLAUDE.md): how you like Claude to work, voice, tools
Vault (your files folder): how your files are organized
Business folder: what the business does, clients, pricing
Project folder: goals, deadlines, people involved
Open anything inside that project and Claude walks up the tree, pulling all four into context. The first message already knows you, your business, and your history.
Pro tip from Taylor: run a "wrap" routine at the end of every session that updates the relevant CLAUDE.md files before you close. Next time you come back, the context is fresh, not stale. Starter repo: claudesidian.
Here's the end-of-session prompt to copy into Claude Code as a /wrap slash command:
Before ending this session, update the CLAUDE.md files at the relevant
folder levels (global, vault, business, project) with any decisions,
preferences, or context from this session that would help the next
session start smarter.
Rules:
- Be surgical. Don't bloat the files.
- Only add information that is likely to still be relevant next session.
- Put each piece of context at the lowest folder level it applies to.
- For each edit, tell me which file you changed and what you added.Our favorite insight: your CLAUDE.md files are the difference between talking to a new employee every session and talking to one who's been with you for a year. Keep them current.
Want more tips like this? Check out our AI Skill of the Day Digest for April.
Total AI beginner? Start here (goes with this video).
Have a specific skill you want to learn? Request it here.
🍪 Treats to Try
Claude Opus 4.7 is Anthropic's new flagship with a 3x vision upgrade (images up to 2,576 pixels), #1 on Vals AI's Vibe Code Benchmark at 71%, and a new xhigh effort level; same $5/$25 per MTok pricing as 4.6, though the new tokenizer can use up to 35% more tokens for identical prompts —$20/mo Pro, $100/mo Max.
OpenAI's overhauled Codex turns the desktop app into a full agent workstation with Mac-level computer use, an in-app browser, automations that wake up across days, and 90+ plugins (Atlassian Rovo, CircleCI, Microsoft Suite) —free with a ChatGPT account.
Canva AI 2.0 lets you describe a project in plain English and get an editable, iterable design with persistent memory, background scheduling, and full orchestration across Canva's suite from one prompt —free tier, Pro from $15/mo.
Alibaba's Qwen3.6-35B-A3B is a new 35B open sparse model (3B active) that rivals Claude Sonnet 4.5 on vision; incontroversial upset, Simon Willison's laptop-local version drew a better pelican than Opus 4.7 Le gasp!—free to try.
Google AI Mode's side-by-side browsing on Chrome desktop opens any AI Mode link as a full page next to the chat, plus searches your open tabs, images, and files —free.
📰 Around the Horn
Factory raised $150M from Khosla at a $1.5B valuation for autonomous coding agents that switch between models by task complexity; Keith Rabois joined the board. The coding agent war has a third front.
OpenAI's chief economist published an AI Jobs Transition Framework mapping 900+ occupations: 18% face higher near-term automation risk, 24% will reorganize (fewer workers, same role), 12% could grow from AI, 46% see less change. ChatGPT is used 3x more in the "at-risk" bucket.
OpenAI launched GPT-Rosalind, its first life-sciences model for biochemistry, genomics, and drug discovery, in research preview with Moderna, Amgen, the Allen Institute, and Thermo Fisher.
The White House is preparing to give federal agencies access to Anthropic's Mythos model, per Bloomberg, while Google is negotiating a classified Gemini deal with the Pentagon, per The Information.
Perplexity released Personal Computer, giving the Mac app the ability to read and write your local files and drive iMessage, Mail, and Calendar (rolling out to Max and the waitlist).
Want absolutely EVERYTHING that happened in AI this week? Click here!
💡 Intelligent Insights
Five think pieces worth your time this week:
Services: The New Software (Sequoia): The next $1T company won't sell software; it'll sell the work. Mapping "autopilot" opportunities across insurance ($140-200B), accounting ($50-80B), tax advisory ($30-35B), and recruitment ($200B+). If you run a SaaS company, this is the thesis eating your lunch from above.
What I learned this week (Dwarkesh Patel): Technical notes on pretraining parallelisms, why distillation probably can't be stopped (agentic tool use happens locally and can't be hidden), and the Mythos cybersecurity equilibrium.
AI cybersecurity is not proof of work (Salvatore Sanfilippo): Finding bugs doesn't scale with GPU; it scales with model intelligence. Weak models pattern-match without understanding, mid-tier models miss the bug, only frontier models chain the reasoning. Pairs with Dwarkesh's notes above.
Why Sal Khan is rethinking how AI will change schools (Chalkbeat): Khanmigo has been "a non-event" because students don't seek it out. Khan Academy is now baking AI into practice problems instead of positioning it as a standalone tutor.
"This is really serious sh$t": OpenAI policy czar thinks 'doomers' are playing with fire (SF Standard): Chris Lehane on the AI backlash and why OpenAI thinks the policy narrative is tilting too far toward precaution.
A Cat’s Commentary
 | That’s all for now.
|
P.S: Before you go… have you subscribed to our YouTube Channel? If not, can you?
Click the image to subscribe!
