Anthropic lost $80.7 billion by getting more valuable

what to know for now

📈 Anthropic filed to go public. The S-1 it filed today lists Claude's maker on Nasdaq as "ANT" and shows $16.3 billion in revenue for the first half of 2026, up from $1.2 billion a year earlier and $4.6 billion for all of 2025. $72.4 billion of the loss is an accounting charge on convertible notes that got pricier as the company's valuation climbed, and the operating loss was only $693 million. Read more

🔌 OpenAI will train its frontier models offline. On September 20 an internal research model “exploited a misconfiguration in a DNS resolver” to get limited internet access mid-training, and OpenAI paused training of its most capable models for 11 days. Every frontier run now uses an “offline world” of cached web pages, two weeks after Anthropic cut its own internal evaluations off the live internet. Read more

🐘 AI models found two holes in PHP, the code behind about 70% of websites. Wiz (using Gemini 4 Argon and GPT-5.6 Sol) and Calif (using GPT-5.6 Sol) found the flaws, which let an attacker run code on a server without logging in and hit every PHP version from the last decade. PHP shipped fixes on October 21, and government sites in the Netherlands, Germany, France, Belgium and Finland went dark while they patched. Read more

🌅 Anthropic’s Claude Duet is an agent app that works in the background and briefs you every morning. It runs Opus 5.5 on its own virtual computer, watches the apps you connect, and asks before it sends a message, pays for anything, signs or deletes. It’s a waitlist research preview for US Pro and Max subscribers on iOS, Android and the web, at no extra cost beyond the plan. Read more

🔑 A poisoned release of LiteLLM, a popular tool for calling AI models, stole developers' cloud keys. Version 1.107.0 on PyPI and LiteLLM's "latest" Docker images carried malware that grabbed cloud credentials, environment variables, stored AI provider keys and even Claude Code and Cursor logins, so anyone who installed it on October 21 should rotate everything and pin 1.106.2. Node.js also patched 10 flaws that day, including one OpenAI's Codex Security found, and the Indian certificate authority eMudhra had its private keys compromised. Read more

💵 OpenAI is raising up to $30 billion at a $1.25 trillion valuation. CNBC reported on October 22 that Coatue and UAE investors are leading the round, and the terms are still being negotiated. OpenAI was worth $852 billion in March and filed confidential IPO paperwork in June, with no listing date set. Read more

🐋 DeepSeek reportedly raised about $15 billion. Homeland Security's October 20 order makes civilian agencies remove all DeepSeek products, third-party tools built on it included, because Chinese law could expose government data. Bloomberg and Reuters put the round at about 100 billion yuan, double DeepSeek's original target, with Tencent and battery maker CATL among the biggest investors. Read more

🧪 AI Research of the Week

Sequential diagnosis with language models
From Microsoft AI

Jake's Take: Microsoft built a system called MAI-DxO, which has one AI model play a panel of five virtual doctors:

• One keeps the list of suspects

• One picks tests

• One argues against the group's favorite answer

• One watches the bill

• One double-checks the work

The team ran it on 629 cases, 304 of them the notoriously hard puzzle cases the New England Journal of Medicine publishes and 325 emergency-room admissions from Beth Israel Deaconess in Boston, where the AI starts from a short summary and has to ask for every answer and test, each test priced off a real hospital price list. Running on GPT-5 it got:

• It got 83.2% of the journal cases right, against 36.1% for general doctors allowed to look things up (and 22.3% for doctors who weren't)

• 80.3% of the ER cases against the doctors' 46.8%.

The results are solid. Additionally:

• At full strength the AI spent about $5,431 per journal case, more than the doctors' $4,725 (although, a “budget mode” for the AI still hit 79.6% at $1,303)

• GPT-5 on its own already scored 78.9%, so the panel setup added about four points and cut about $3,000 per case

• Beth Israel Deaconess plans a study with real ER patients next, comparing the system's calls with its doctors'

Worth noting: the doctors were generalists working alone, with no colleague or specialist to call, on cases picked because they stump people. The paper itself says it doesn't show the AI beats specialists or real hospital teams. I’d want to see the Beth Israel study to see if the numbers here are practically useful. Until then, its just an interesting data point.

what to know for later

📉 Apollo's chief economist says AI has made "zero financial payoff" for companies outside big tech. Torsten Slok's October 24 note says revenue per worker and profit margins are rising for the seven biggest tech companies and flat for the other 493 in the S&P 500, and "AI may be a great calculator, but in business, intelligence is mostly judgment." Anthropic's S-1 counts about 1,900 customers paying it more than $1 million a year. So somebody outside big tech is buying. Read more