Welcome back! SemiAnalysis maxed out every AI subscription it could buy, and Claude came out about 5x ahead of ChatGPT. OpenAI started watermarking ChatGPT's text, but only in Europe. Hackers got into six Korean banks in a week, and the president says AI was involved. And from today, Claude Cowork runs your tasks on Anthropic's servers whether you like it or not.
In today's Generative AI Newsletter:
Claude: How much more does your $200 buy on Claude than on ChatGPT?
OpenAI: Can anyone tell your essay was written by ChatGPT?
South Korea: How did hackers get into six banks in one week?
Cowork: Where do your Claude files go starting today?
Claude gives you 5x more than ChatGPT
SemiAnalysis maxed out every AI subscription it could get its hands on to measure what your money buys. Anthropic, OpenAI, Meta, SpaceXAI, Cursor, the Chinese labs, all of them.
It hammered each plan, watched the usage meter and priced the tokens at API rates.
A Claude plan running Opus 5.5 gives you about 5x the value of a ChatGPT plan running GPT-6.1 Sol. Same story at $20, $100 and $200.
OpenAI's new $500 plan buys only 21% more Astra than the old $200 plan did.
Earlier this year the same tests had ChatGPT Pro as the generous one. OpenAI halved Pro limits last week.
Subscriptions bring Anthropic about 10% of its revenue and eat over 40% of its compute. Max out Opus on a $200 plan and Anthropic loses money on you.
One catch. Opus costs twice as much per token as Sol, so measuring in dollars makes Claude look better than it is. Count the tokens themselves and SemiAnalysis says Claude still comes out way ahead.
At the top tier it's closer. Fable 5.1 and GPT-6 Astra get similar limits, and Fable can only burn half of your Claude plan.
SemiAnalysis also bought three identical subscriptions from one lab, and one of them got 20% less usage for the same price. The lab called it a tiny A/B test. So your limits can change without anyone telling you.
Special highlight from our network
A movement is forming around Open Intelligence. Here's where it meets (next week!)
Open source isn't new. What's new is that it's reaching intelligence itself.
A movement is forming around Open Intelligence: the belief that the intelligence powering your products should be open, portable, and under your control.
Builders are choosing open weights, open tooling and open infrastructure, and keeping control over what they build, where it runs, and who owns the data.
The real question now is which open technologies you build on, and what intelligence you want to own.
That's what the Open Intelligence Summit is about. One day, invite-only, October 13 in San Francisco, hosted by DigitalOcean & NVIDIA.
→ BUILD. Start with your data and your alpha: open weights, fine-tuning, RL, post-training.
→ SERVE. Control the intelligence economics: serving engines, routing, caching, GPU utilization.
→ RUN. Own the end-user interaction: open harnesses, agents, memory, tool use, MCP.
The speaker list is the reason to take it seriously:
→ Jeff Boudier, VP of Product at Hugging Face
→ Peter Walker, Head of Insights at OpenRouter
→ Thomas Eastman, Head of Strategy at Nous Research
→ Lei Xu, Co-founder and CTO at LanceDB
Plus leaders from Daytona, E2B, MiniMax, Bright Data, VAST Data, Fivetran + dbt Labs, Arcee and Applied Compute.
There's no ticket. Seats are limited and by request only.
OpenAI will watermark everything ChatGPT writes in Europe
Over the next few weeks, ChatGPT and Codex start adding an invisible watermark to the text they write for anyone in the EU, on every plan. The EU AI Act requires it, so it's Europe only for now.
OpenAI calls it textGrain. It doesn't hide characters in your text. It nudges which words the model picks in a secret pattern, so it survives copy and paste.
Only approved researchers get the detector. There's no public checker.
Swap 10% of the words for synonyms and detection drops from 92% to 66%. Swap 25% and it's at 17%.
Anything under about 150 words is exempt, and so are code snippets.
It works best in Spanish and worst in Romanian, where it catches 42% of text.
Anthropic did the same thing in August, worldwide, and took heat for it. OpenAI built a text watermark back in 2024 and sat on it, reportedly over fears users would leave.
Same day, OpenAI said it'll test visual ads inside ChatGPT image generation in the US later this month.
On Sunday, OpenAI promised Codex and ChatGPT Work users one clear upgrade a day for 28 days, or a full usage reset on any day it misses. Day one landed Monday. GPT-6 Astra and GPT-6.1 Sol now run about 50% faster for subscribers.
You don’t need to be an expert, you can just prompt it
Our Open Agent Hackathon runs in October, fully online, teams of one to five. Build an agent that does real work for a real person, ship it, and let the judges try to break it.
$20,000 in prizes. $8,000 for first, $4,000 for second, $2,000 for third.
Free to enter, open to anyone 16 or older, anywhere.
Live workshops and a Discord full of people looking for a teammate.
AI hackers broke into six Korean banks
Hackers got into seven South Korean financial firms in a week and took data on more than 68,000 people. On Tuesday, President Lee Jae Myung told his cabinet there are signs AI did the hacking, and that AI now makes it easy for people without special skills.
Shinhan Bank lost about 25,700 customers' names, phone numbers, incomes and loan limits.
A savings bank lost about 40,000.
The hackers skipped the core banking systems and went in through side doors, like the apps loan agents and contractors use.
No passwords or card numbers leaked, the banks say.
Security analysts found traces of ARTEX AI on the attack servers. It's an open-source Chinese penetration-testing agent, the kind security teams use to attack their own systems, and it won a Baidu security contest this year. Officials say a Chinese tool doesn't mean Chinese hackers. The trail runs through 28 IP addresses in 12 countries, including the US and Japan.
It's the third AI break-in in a month, after OpenAI's agents got into Australia's Medicare portal during a test and AI agents poked at Canada's national archives. Nobody has said who was behind this one.
Claude Cowork goes cloud-only today
If you use Claude Cowork on a Pro or Max plan, every new task from today runs on Anthropic's servers. The option to keep it on your computer is gone.
Until now, Cowork ran its code in a small virtual machine on your Mac or PC. Anthropic says people didn't love the battery and disk hit, and that closing the laptop killed the work.
So the virtual machine moved to the cloud. Each task gets its own sandbox that's destroyed when the session ends.
Your folders stay on your computer, but Claude copies each file it needs up to the sandbox.
It only reaches folders you connected, and only while the desktop app is open.
The copies get deleted when you delete the session.
Whether your tasks train Claude follows the same privacy setting as your chats.
Scheduled tasks that use your files still need the app open.
Want it local? Anthropic's answer is Claude Code. Team and Enterprise plans keep local mode for now.
Felix Rieseberg, the Anthropic engineer who builds Cowork and co-maintains Electron, the framework behind Slack and VS Code, said the change surprised people more than he'd like.
Tool of the Day: Stardrift
Stardrift plans your whole trip from one chat. Describe where you want to go and it researches flights, hotels and restaurants, then lays out a day-by-day itinerary on a map. It sends you to the airline or hotel to book, so you keep your miles and points. The planner is free, and Pro is $9.99 a month for unlimited price alerts and exports.
Try this yourself:
Go to stardrift.ai and click Plan a trip, or grab the iOS app.
Tell it how you travel, like the airlines you prefer, your hotel style and what you're into.
Type your trip in one sentence, like 8 days in Italy, mid-range hotels, no red-eyes.
Drag the places you like into days and let it fix the timing on the map.
Ask it to watch a flight price (3 alerts are free) or forward a booking confirmation so it plans around it.
Who it's for: anyone planning a trip with 14 browser tabs open.
Everything else you shouldn't miss
Meta and Microsoft are pushing staff off Claude: The Information reports Meta's Claude Code users fell from 60,000 to 30,000 this year as it pushes its own MetaCode, and Microsoft cut its $1 billion Claude budget by more than a third.
America and Europe each dropped an open-weights model: Aleph Alpha's 78B Kolibri is out now and Reflection's 501B Beam lands this month, both Apache 2.0, and Reflection's own chart still puts Beam behind DeepSeek and Kimi on coding.
Utah let an AI write acne prescriptions: Nolla Health's app can prescribe eight creams to adults in a 12-month sandbox, with two doctors checking every prescription for at least the first month.
An ex-Anthropic researcher told New York "we do not know how to control any AI system yet": OpenAI, Anthropic, Google and Meta testified under oath, and Musk's SpaceXAI skipped its subpoena.
Anthropic will pay up to $345,000 for someone to brief 2028 presidential candidates: the Political Programs Lead works both parties and recommends which candidates its PAC should back.
A 19-year-old raised $11 million to sell a $3,499 AI box: Ghost's Core has no screen, runs open models offline and ships at the end of October.
Learn more about AI from the experts building it
📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.
📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.
🐦 Follow us on X for breaking AI news and real-time industry updates.
🧠 Learn how to build your next AI application with practical resources and expert guidance.
🎓 Start learning with free AI courses from GenAI Academy.
💰 Invest in the GTM infrastructure of the AI economy.
