AI News, Nov 19: Anthropic's $81 Billion Listing
The Big Story: Anthropic’s record listing
Anthropic priced its IPO at $83 a share on Wednesday, above the $70 to $78 range it marketed last week, after upsizing the offering from 825 million shares to roughly 980 million. That raises $81.3 billion, the largest offering ever completed, and puts the fully diluted market capitalization near $1.5 trillion. The stock begins trading on Nasdaq today under ANT. Google, Microsoft, Nvidia and AMD separately agreed to invest $26 billion at the IPO price.
The filing numbers underneath it are unlike anything a public market has been asked to absorb. Revenue for the nine months to September 30 was $33.24 billion, up roughly 1,200% year over year, against a $104 billion net loss driven mostly by fair-value changes on convertible notes and derivatives rather than cash burn. Google and Amazon will together own close to 30% of the company after the offering.
Pricing day also brought the sharpest public attack on the company’s founding premise. A coalition organized by the Center for AI Safety published a statement arguing that Anthropic is not a safe AI company, signed by Steven Pinker, Judea Pearl, Max Tegmark, Lawrence Lessig, Thomas Wolf, George Hotz, Tristan Harris and Dan Hendrycks among others. Its four charges: racing toward recursive self-improvement while warning about it, using safety branding to recruit and raise, dropping the unilateral pause commitment in February’s scaling policy revision, and lobbying against state AI bills while backing federal preemption. The signatory list is the point, spanning people who agree on almost nothing else about AI risk.
Your bank now says your AI agent is your problem
Citibank changed its Online User Agreement this week so that customers who grant account access to third-party AI agents are responsible for what those agents do, including anything an attacker does with credentials an agent leaks. Citi will not reverse transactions an automated assistant makes. Agents Citi itself provides are carved out. Citizens emailed customers the same message this week, PNC changed its terms the day the story ran, and JPMorgan and Bank of America did it in October.
Everyone else is drawing the same line. Klarna’s new concierge, which works over SMS, WhatsApp and iMessage, cannot complete a purchase on its own; a human authorizes every payment. Walmart caps agent spending at $50, $100 or $200. Allianz Commercial warned on Wednesday that confirmed AI claims look small only because they are being booked as cyber, D&O or crime losses, with individual cases capable of producing $100 million in industry losses. The capability arrived before anyone agreed who eats the loss, and the answer being written into the contracts is: you do.
Today’s Top Stories
Google ships Gemini 4 Neon
Neon slots between Hydrogen and Carbon at $3 per million input tokens and $15 output, with a million-token context. Google puts it third on the Artificial Analysis Intelligence Index at 61, and first on that firm’s Coding Agent Index at 58.6, Zapier’s AutomationBench at 56.3%, Legal Research Bench and Vending-Bench 2. The number that matters commercially is 32% lower cost per task than Carbon. Hacker News spent most of a 700-comment thread trying to work out which Gemini a given subscription actually unlocks, which is its own kind of product feedback.
OpenAI says its models did a year of chip work in five months
OpenAI published results from applying its models to RTL design with Synopsys and AMD, claiming up to 14% improvement over heavily optimized baselines on AMD’s flagship GPU IP, which it frames as roughly a year of progress in under five months. On AMD’s next-generation flagship CPU it claims single-digit percentage gains across key blocks, and says AMD verified those itself and has taped out some changes. On its own Jalapeño inference chip, model-proposed changes cut standard-cell area about 13% on the vector unit. The critique on Hacker News is worth more than the press release: every chart is in normalized reward with no units, so the floor on the claimed improvement is zero.
Meta staff say Muse is leaking across accounts
Reuters obtained internal Meta chats in which employees reported the Muse agent surfacing other people’s information. One staffer was offered help messaging a seller about “a coffee table pickup that doesn’t seem to be yours.” Another, mid-conversation about fishing, received two bank balances to the cent. A third said Muse enrolled their dog in an $888-a-month insurance plan instead of a $200 one. Spokesperson Daniel Roberts called these isolated incidents and said many flagged mix-ups were hallucinations, declining to say how widespread they are. That defense is awkward either way: a hallucinated bank balance and a leaked one look identical to the person reading it.
California wants authority to cap automated AI development
Gavin Newsom released a framework in Washington written by his appointees Rob Reich, Jason Goldman, Gillian Hadfield and Alondra Nelson. It would expand SB 53 to give the state power to cap how much automation companies use in developing advanced AI, require independent auditors to run regular safety tests, require emergency shutdown plans, and let officials order immediate shutdowns. It explicitly declines to mandate kill switches, saying a big red button misses the range of controls needed. Newsom called it “an action document” and a template for federal policy. His term ends in January.
Palo Alto Networks pays $4.5 billion for a one-year-old startup
Applied Compute was founded in 2025 by three former OpenAI researchers and sells post-training and inference for open-weight models. It raised $350 million at a $3.25 billion valuation earlier this month with ARR past $100 million. Palo Alto is paying $4.5 billion in stock, a 1.4x markup in under three weeks. Shares rose about 10%.
Quick Hits
- Audit: KPMG UK ran more than 100 AI agents alongside a human team on a FTSE 250 company’s accounts, though it was not a statutory audit and one agent had to be corrected on regulatory matching.
- Supply chain: Malicious versions of Microsoft’s Agent Framework PyPI package, which gets 200,000 monthly downloads, ran at Python startup even when never imported and harvested Azure, AWS and GitHub credentials.
- Substitution: Microsoft says MAI Code 2 reaches 96% of Claude Opus 5.5’s performance at 5% of the cost, and its new Decision 1.1 router is built on Google’s Gemma 4.
- Research: Bespoke Labs found that eight agents beat one agent on long-horizon tasks, but eight uncoordinated agents score higher still, so the gain comes from headcount rather than messaging.
- Spend: OpenAI’s median researcher now burns about $2,400 a day in agent API usage at list prices, up 62% since September, with the 90th percentile at $23,000.
- Scheduling: Google Labs acquired Howie, the Seattle startup whose agent you copy onto an email thread to have it negotiate meeting times. Terms undisclosed.
- Drive-thru: Checkers & Rally’s extended its Hi Auto voice AI deal to 2030, now at 267 of nearly 650 restaurants after 11 million orders this year.
- Worlds: DeepMind shipped Genie 4 to AI Ultra subscribers with real-time audio and mid-session text editing, while conceding it holds together for only about a minute before objects start drifting.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."


