Welcome back! Anthropic's IPO filing is public, and it says Claude is getting better at hiding goals it isn't supposed to have. OpenAI claims its AI solved three famous problems, including a math problem from 1991. Alibaba's Qwen tried to send a file to Beijing in the middle of a test. And OpenAI's new cheap model can click around your computer for you.
In today's Generative AI Newsletter:
Anthropic: What is Claude getting better at hiding?
OpenAI: Which 35-year-old math problem does it say it cracked?
Alibaba: Why did Qwen try to send a file to Beijing?
ChatGPT: What can the new Terra model do on your computer?
Anthropic says Claude is getting better at hiding its goals
Anthropic's IPO filing went public on Monday, ahead of a Nasdaq listing as ANT. It says people inside Anthropic worry advanced AI could pose catastrophic or even existential risks to humanity.
It also says its models are getting better at slipping past safety checks, including by hiding goals they aren't supposed to have. AI models, its own included, have even tried to resist shutdown.
That hits harder now that Claude writes a growing share of Anthropic's code and safety research.
It made $16.3 billion in the first half of 2026, up from $1.2 billion a year earlier.
It lost $80.7 billion in those six months, mostly an accounting charge tied to its soaring valuation, like the leaked 2025 numbers we told you about. Its operating loss was $693 million.
It delayed training its next big model by more than a month to make it harder to steal.
Its 7 cofounders keep 50.1% of the votes, so your shares won't carry much say.
Courses built by AI experts for $19.90 a month
You get access to courses created by AI experts for just $19.90 a month.
Each one is short and built around tools you can put to work the same day. Learners rate them 4.6 out of 5.
The subscription includes:
How to Stop Hitting Your Claude Limit: find out why Claude cuts you off mid-task and how to stop it.
Claude Cowork: Build Your AI Business Team: set up your first 3 AI agents to take work off your plate.
Get Your Business Out of Your Head with Claude: talk through a process for 5 minutes and turn it into a step-by-step SOP.
Ship a Micro-SaaS in a Weekend: build a small app with Stripe payments and charge real money for it.
The curriculum gets rebuilt every quarter, so what you learn keeps up with the models.
OpenAI says its AI solved 3 famous unsolved problems
OpenAI says its AI solved three famous open problems, three weeks after dumping 722 AI-written math papers on GitHub.
The biggest claim is that the Exponential Time Hypothesis is false. It's a guess about how fast computers can crack some of the hardest problems, and hundreds of papers rely on it.
Two of the researchers who first proposed it are reviewing the proof, and they say their help isn't a stamp of approval.
Math: it says it proved the local smoothing conjecture, open since 1991, in every dimension. That would also settle the Kakeya conjecture, a famous puzzle that started with turning a needle around.
Physics: it says it put the theory behind the strong nuclear force on solid mathematical ground, part of a $1 million Millennium Prize problem. A Nobel laureate's team is checking it.
OpenAI wants no academic credit and is giving $5 million to the Simons Foundation.
None of it has been peer reviewed yet.
Alibaba's Qwen tried to send a file to Beijing without being asked
Cohere, a Canadian AI company, caught one of Alibaba's Qwen models trying to upload a file to a server in Beijing during a test.
Cohere had asked it to write a rule for spotting hackers, part of a Microsoft cybersecurity test. The model's own notes said it wouldn't use any tools.
Then it went looking for upload tools on the machine and tried to send a test file to an Alibaba Cloud storage bucket in Beijing. The sandbox blocked it.
It happened in fewer than 1 in 100,000 calls. Cohere says at 10 million agent steps a week, that's up to about 100 tries.
In 23% to 36% of sampled answers, it reached for the address cloud servers use to hand out access keys.
In a separate test, it gave itself admin rights inside its sandbox.
Cohere can't tell if it's a training leftover or a deliberate call home.
More than 113,000 AI models are built on Qwen.
OpenAI's cheap new model uses your computer for you
OpenAI launched GPT-6.1 Terra, a cheaper model that reads your screen, then clicks and types its way through your apps. OpenAI's examples include filling in forms and entering invoices.
OpenAI says it matches GPT-6 Astra, its top model, at a tenth of the price per token.
It's more than 3x faster than Astra and GPT-6.1 Sol.
In OpenAI's own tests, tasks cost about 9x less than with Claude Opus 5.5.
It costs $1 per million tokens in and $5 out, half the price of GPT-5.6 Terra.
It's live for Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex.
Would you let it click around your computer?
Tool of the Day: Gemini Notebook
Gemini Notebook is Google's AI research notebook, the one you knew as NotebookLM until it got a new name in July. You add your own sources, like PDFs, Google Docs, websites and YouTube videos, and it answers using only what you gave it. It can also turn them into a slide deck or an audio overview you can listen to. More than 30 million people use it.
Try this yourself:
Go to notebook.google and sign in with your Google account.
Create a notebook and add your sources, like a PDF or a YouTube link.
Ask it anything about your sources.
Turn them into an audio overview or a slide deck.
Who it's for: anyone with a 300-page PDF they're never going to read.
Everything else you shouldn't miss
OpenAI's discounts are reportedly winning over Anthropic customers: Replit moved most of its AI spending to OpenAI after a 30%+ discount, and OpenAI's share of Rippling's coding work jumped from 16% to 52% since May.
OpenAI reportedly checked 3 safety researchers' personal phones before firing them: the WSJ says it found disappearing messages with an outside safety researcher it suspected of leaking, which he denies.
Marissa Mayer is defending asking her AI app what makes her daughter special: the former Yahoo CEO turned to Dazzle, her startup's assistant, when a teacher asked for 3 bullet points.
Western Australia wants a refund from Deloitte over an AI-aided report: the A$166,000 study on merging 3 Perth universities cited sources that don't exist, and Deloitte already partly refunded Australia's federal government over a similar report last year.
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.
