AI News, Nov 20: Anthropic's Debut Lasted One Day
The Big Story: a record debut with a one-day shelf life
Anthropic’s first session on Nasdaq went about as well as a listing can go. Shares popped 34% above the $83 IPO price at the open and closed at $97.17, up 17.07%, with more than 640 million shares traded by mid-day, beating SpaceX’s entire opening day in June. The close valued Anthropic at roughly $1.8 trillion fully diluted, with Amazon owning 14.7% of it and Google 13.7%. The underwriting banks collected $550 million.
Then Friday. Anthropic fell in pre-market trading, at one point down over 6%, with Nvidia and AMD lower alongside it. SoftBank had its worst day in over two years as the selloff swept Asian tech, and Japan’s Nikkei 225 and South Korea’s Kospi each plunged nearly 5% intraday, led by chipmakers. The largest offering ever completed did not steady the AI trade; it gave the market a liquid way to bet against it.
More than 100 protesters gathered outside Anthropic’s San Francisco offices chanting “ban AI that builds AI,” and Rep. Pramila Jayapal said on X she was “alarmed that Anthropic is barreling ahead with its IPO today despite a series of outages in critically important systems.” Dario Amodei, from the Nasdaq podium: the listing “only strengthens our resolve to do the right thing.”
Everyone shipped a computer-using agent, and everyone else started blocking the screen
Google switched on Computer Use in the Gemini Mac app, which can now see the screen through screenshots and “click, type, scroll and navigate programs and websites for you.” OpenAI pushed Dots, its always-on agents that can connect to your PC and control it on your behalf, down to the $20 Plus tier and out to every ChatGPT Business Standard user. Apple is hiring a manager to build an agent that “can use a computer the way a capable person does”.
The layer underneath moved the other way inside the same day. Apple’s macOS 27.3 beta adds granular control over what third-party AI agents can reach. Signal turned screenshot blocking on by default on iOS 27 to keep system assistants out of chats, admits it cannot do the same on macOS, and wants Apple, Google and Microsoft to publish a way for apps to opt out of AI access entirely. Banks got there first, largely closing their doors to outside AI agents even as TD Bank’s US chief predicts “every client will have a dedicated virtual assistant.” Permission to see the screen is becoming the scarce input, not the agent.
Today’s Top Stories
Intuit sold Mailchimp for less than a third of what it paid
Intuit is handing Mailchimp to Bending Spoons for about $3.68 billion, five years after paying $12 billion for it in what was then the largest acquisition in company history. Mailchimp revenue fell 1% to $313 million in the quarter ended Oct. 31. Intuit reinvests $368 million of the proceeds in Bending Spoons shares and keeps $3.31 billion in cash; its own stock fell more than 10% after hours anyway. Bending Spoons has now absorbed Airtable, Miro and Mailchimp since listing in July, making it the decade’s busiest buyer of software nobody else wants to operate.
Washington put the China chip trade on the record, and AMD paid for it
California Attorney General Rob Bonta subpoenaed Nvidia over whether its chips reached China against export restrictions and whether investors were misled: “The conduct of Nvidia and its former CEO raises bright red flags.” The House Foreign Affairs Committee separately released the licensing records, showing Commerce approved 879,000 AI chips for China from December 2025 to May 2026 out of 3.4 million requested, 150,000 of them for Tencent. AMD then closed down 7.72% at $489.19 after Hunterbrook Media, disclosing a short position, alleged it “bills over 20% of its revenue to customers in China and Hong Kong.” AMD said the claims “misrepresent” its business practices and relationships.
Frontier models can name who wrote a text, from the text alone
Given one piece of writing, no web access and no candidate list, frontier models match trained AI-text detectors and then go past them. Opus 5.5 names the correct originating lab for 98% of texts generated by Anthropic, OpenAI and Google models. Fable 5.2 identifies the author of 39% of public peer reviews published in October and November, rising to 78% on a subset of Nature reviews, and the username behind 69% of comments from Redditors with more than 500 posts in the past year. Accuracy tracks how much of an author’s writing sits in pretraining, and the authors report defenses are limited, which makes anonymous peer review a thinner promise than it was last week.
The AI spending curve bent, in three datasets at once
Ramp, tracking more than 85,000 US companies, found tokens 21.2% cheaper in six weeks and volume at a record high, but spending down 6.3% from its peak two weeks earlier. Stripe, across roughly 20,000 AI companies, put revenue growth at 103% over the 52 weeks to Nov. 8 while the median AI company grew 41%, and 13-week annualized growth at about 80%, down from 140% a few months ago. On the buying side, KPMG’s survey of 100 insurance chief executives found nearly two-thirds planning to put 11% to 20% of total budget into AI, against 1% reporting a measurable return.
Quick Hits
- Security: the European Commission is into a second week of disruption from an attack built on the public ARTEX offensive-AI framework, with the FBI counting 1,600-plus ARTEX consoles exposed online and attackers reaching privileged access in under three hours.
- Open weights: Tencent released Hy4, a 770-billion-parameter multimodal model, then dropped the preview’s Apache 2.0 licence for one requiring written permission from hosted-inference businesses above $50 million in revenue or 100 million monthly users.
- Breach: the Fireworks AI compromise widened, with Notion saying transcripts from under 1% of workspaces were exposed and disconnecting, and Uber reporting stolen driver licence data.
- Outages: officials told a House hearing there was “no indication whatsoever” of malicious activity behind the failures at Microsoft, the FAA, JPMorgan and the NSA, blaming the scramble to patch what frontier models found.
- Policy: Reps. Beyer, Jacobs and Maloy outlined a bipartisan SAFE AI Act mandating safety frameworks, incident reporting and transparency from frontier labs, with civil penalties and emergency authority to halt dangerous models.
- Retail: Target said Ask Target beta users purchased at about twice the rate seen elsewhere and Walmart put Sparky usage up 129% year over year, with neither naming a model vendor.
- Cost control: Uber overran its annual AI budget in months, then cut cost per 1,000 agentic requests by almost 34% partly by loading no tool definitions at startup, since 100-plus tools burn 50,000 tokens before anyone types.
- Dissent: more than 500 Google employees and the Alphabet Workers Union broke with the company over bundling AI into Search for schools, asking that schools be allowed to switch it off.
- Energy: the IEA expects data centre electricity use to nearly double within four years, with AI driving about 75% of the added capacity and emissions at risk of climbing 70%.
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."

