Anthropic, the company behind Claude, filed to list on the Nasdaq as ANT, with trading expected in mid-November, and the filing shows both sides of the AI boom. Revenue reached $16.3 billion in the first half of 2026, 14 times as much as a year earlier.

Yet Anthropic has lost money every year since it was founded: $42 billion last year and $80.7 billion in the first six months of this one. Most of that is an accounting charge on its convertible notes, and running the business lost $693 million, but the bills keep growing: it has agreed to spend up to $120.5 billion on computing from SpaceX's AI unit, formerly xAI, alone. Blackstone, which leads a $63 billion deal to pay for Anthropic's chips, now expects its debt to be rated junk after the listing, TechCrunch reports.

Investors spent the week asking when all that spending pays off. Alphabet's shares fell 12.6% on Thursday, their worst day ever, wiping out about $500 billion after a weak quarter for search ads, which a Google executive partly put down to its AI search mode showing fewer of them. On Friday tech stocks lost more than $1 trillion between them, as Amazon said it would spend "meaningfully more" next year and Intel warned of weaker sales, NBC News reports.

OpenAI says its AI disproved a famous assumption

OpenAI announced three discoveries in mathematics, computer science and physics. The one that drew most attention is a 67-page paper in which, OpenAI says, its AI models disprove the Exponential Time Hypothesis, an assumption computer scientists rely on about how hard a classic kind of logic puzzle is to solve. Three leading computer scientists are checking the proof and wrote that their involvement "should not be taken as certification of the claim", Decrypt reports. It doesn't settle the bigger P versus NP question, and OpenAI says turning it into anything practical "will require significant further research". On Hacker News, where the announcement and the paper drew close to 2,000 points between them, one of the top comments credited OpenAI for bringing in mathematicians before announcing this time.

Google's new top model goes to security teams first

Google released Gemini 4 Carbon, its new top model, and claims it's the best yet at finding and fixing security holes. For now only selected security teams can use it, and Google gives no date for everyone else. Its launch post drew about 34,000 likes on X, but on Hacker News the top comments complained that Google keeps announcing models long before people can use them.

Mistral gives away its biggest model

The French lab Mistral released Mistral Large 4 for anyone to download and run, businesses included; only companies that resell it as a service need a separate licence. It's far too big for a laptop, so most people will use it through Mistral's own service or others that host it. It drew 1,700 points on Hacker News, where one of the top comments read: "The more countries are in this game the better for us all."

A cheaper OpenAI model that works your computer

OpenAI's GPT-6.1 Terra is built to use a computer for you, filling in forms and clicking through apps. Paid ChatGPT plans get it in Codex and ChatGPT Work, and for developers it costs half what OpenAI charged before, BleepingComputer reports. In a demo by someone who works on ChatGPT and Codex at OpenAI, it filled in a long web form in 53.5 seconds.

Genie 4 builds worlds you can change by typing

Google DeepMind showed Genie 4, a model that generates a scene you can move through and change with typed instructions, now with sound. Only trusted testers can try it so far, and Google AI Ultra subscribers are next. On Hacker News, commenters pointed out details that change when something leaves the frame and comes back.

Chatbots lose to an old model at spotting skin cancer

A London dermatologist tested GPT-6 Astra and Claude Fable 5.1 on 200 skin lesions confirmed by biopsy. Choosing from six diagnoses, both named the right one 42% of the time; a 10-year-old image model trained for the job got 62%. If a mole worries you, show it to a doctor rather than a chatbot.

Fixing devices instead of replacing them

Most of these fixes were a setting or a small program, not a new part.

Homemade 3D lessons

Each of these interactive lessons was built by one person, some in a few days. The careful makers check every number before anyone learns from it.

The Heart in Twelve Leads, a 27-chapter EKG course a psychiatry resident built with Claude Code.

Games made by one person

Of the week's 31 game projects, 19 were built with Claude Opus 5.5, many of them with the AI working directly in Unreal Engine, Roblox or Blender.

Models

Apps

For creators

Unslop is a free Claude skill that reworks landing pages, dashboards and app screens that look AI-made into something that looks designed. It had more than 3,000 stars on GitHub two days after it went up.

Open Source Canvas is a free canvas for making AI video with models that run on your own computer through ComfyUI, modelled on Higgsfield's canvas but with no subscription. Its maker released it under the MIT licence on 29 October.

For developers

Mainplane is a free, open-source program that lets an AI model of your choice operate your computers, using their real browsers and apps, with controls over what it may touch. It runs on macOS, Windows or Linux, on one machine or across several, and its maker introduced it on X this week.

GEOInflow is free, open-source website analytics that shows which AI companies' crawlers read your site, which pages they skip, and how many visitors AI search sends you. It runs on your own server.