← Home AI in 15

AI in 15 — November 19, 2026

November 19, 2026 · 15m 51s
Kate

Eighty-three dollars a share, nine hundred and eighty million shares, and about seventy-nine billion dollars raised for one company. Anthropic priced its IPO last night, above the top of its range, and this morning anyone with a brokerage account can buy a piece of a frontier AI lab.

Kate

Welcome to AI in 15 for Thursday, November 19, 2026. I'm Kate, your host.

Marcus

And I'm Marcus, your co-host. Morning, Kate. We've been counting down to this one all week, and now it's here.

Kate

Here's what's coming up. Anthropic prices its IPO, and trading starts today.

Kate

Google launches Gemini 4 Neon, its new mid-tier model, and goes after cost per task.

Kate

Microsoft's Ignite brings a whole family of in-house models, plus a new chip.

Kate

OpenAI says its models helped design its own chip.

Kate

Plus Genie 4 worlds you can redirect mid-game, a hundred AI agents on a real audit, a tiny Sudoku solver, and a Berlin website owner who says AI is putting him out of business.

Kate

Marcus, on Monday the range was seventy to seventy-eight dollars on eight hundred and twenty-five million shares. What did we actually get?

Marcus

More of both. The price is eighty-three dollars per Class A share, five dollars above the top of the range. The offering went up to nine hundred and eighty million shares. Anthropic sells about nine hundred and fifty-six million of those, and existing stockholders sell the rest, roughly twenty-four million. The underwriters, led by Goldman Sachs and Morgan Stanley, also have a thirty-day option on another hundred and forty-seven million shares. It lists on the Nasdaq Global Select Market under the ticker A-N-T. Trading is expected today, and the deal closes tomorrow.

Kate

And the money?

Marcus

Anthropic's release doesn't give a proceeds figure, so this is our own arithmetic. Its share comes to about seventy-nine point four billion dollars before fees, and the whole offering to about eighty-one billion. If the underwriters take up their option, add roughly twelve billion more.

Kate

So the price went up and so did the number of shares. That usually means a lot of demand.

Marcus

That's the usual reading. When bankers can raise the price and the size together, the order book was oversubscribed, meaning more buyers than shares. But I'd be careful with the valuation. You'll see one point five trillion dollars all over social media today. That comes from a Hacker News commenter, and it isn't in the release. It depends on how you count the share classes and the private placement we talked about on Monday. I'll wait for a proper number from the filing or from the first trade.

Kate

Fair. So why does a listener who'll never buy a share care?

Marcus

Because until now, all the big claims about frontier labs, the twelvefold revenue growth and the nearly nine hundred billion dollars of compute commitments, were tested only by insiders in private rounds. From today there's a public price that changes every second. If that price holds, every lab, OpenAI included, gets a benchmark for going public. If it drops, all those compute contracts look a lot riskier.

Kate

Founded in 2021, and listed at this scale five years later. That's wild.

Marcus

That was half of the Hacker News reaction, pure amazement. The other half was "at what price?" Both are reasonable.

Kate

And there's been some noise around the listing.

Marcus

Yes. A coalition of scientists, legal scholars and tech people put out a press release titled "Anthropic is Not a Safe AI Company," challenging the safety branding ahead of the IPO. We couldn't read the full release, so I won't describe their arguments. Commenters noted the list has both people who think AI is very dangerous and people who think it's overhyped, which is an unusual mix. Timing a release to an IPO is also a pretty familiar move.

Kate

Quick hits. Google released Gemini 4 Neon, with about six hundred and fifty points on Hacker News. Marcus, where does it fit?

Marcus

It's the middle of the Gemini 4 family. Carbon is the heavyweight for the hardest long projects, Hydrogen is the fast, cheap one, and Neon sits between them. Google is aiming it at software engineering, knowledge work and science, especially long agent tasks that call a lot of tools. It has a million-token context window and four thinking levels. In Antigravity, Google's coding tool, it can act as an orchestrator, a kind of foreman directing several Hydrogen subagents working in parallel.

Kate

And the price?

Marcus

Three dollars per million input tokens and fifteen per million output, with cached input at fifteen cents. Google says that at default settings it costs thirty-two percent less per task than Carbon. That's the pitch: not the smartest model, but the most work per dollar.

Kate

Benchmarks?

Marcus

Google's own reporting puts it number one on the Artificial Analysis Coding Agent Index, number two on the Vals Index, and number three overall on the Artificial Analysis Intelligence Index. That third place is actually the useful bit, because it's an independent ranking and Google isn't claiming the top spot. Google also says zero percent cheating on Terminal-Bench and a zero percent success rate for one prompt-injection benchmark at the highest thinking level. Zero percent is a big claim, and those are Google's numbers.

Kate

What do the people actually using it say?

Marcus

It's mixed, in a way that's encouraging. One commenter called it the first Google model that "feels good for coding." Another said it still trails Opus 5.5 in real use, whatever the benchmarks say. And several pointed out the speed. On OpenRouter it's running at around a hundred and twenty-seven tokens a second, against forty-six for Carbon. When an agent makes hundreds of calls, that adds up.

Kate

There's a security version too?

Marcus

"Neon for cyber," going only to vetted defenders through something Google calls the Fairwind Program. Given the week we've had with AI-found bugs, I'm glad that one isn't on general sale.

Kate

Microsoft. We covered MAI Code 2 yesterday. But Ignite brought a lot more than that. Marcus, what's new?

Marcus

The headline says seven new models, and the page lists eight items, because one is a chip. Beyond the coding models there's MAI Cyber 2 Flash, which Microsoft says beats GPT-6 Astra on a cybersecurity benchmark at under two percent of the cost per task. MAI Image 2.7 is ranked first for image editing on Artificial Analysis. MAI Realtime 1 is a speech-to-speech model starting at one cent per audio minute. And there's Decision 1.1, a router.

Kate

A router for what?

Marcus

Think of a hospital triage nurse. Most questions get handled by a small, cheap model, and only the hard cases get sent up to the expensive specialist. That's where a lot of real savings will come from. Then there's Maia 200, Microsoft's second-generation AI chip, which it says produces one point seven times more tokens per watt.

Kate

So Microsoft now has models, a router and its own silicon. That's a whole stack.

Marcus

It is, and for a company whose AI strategy was basically "OpenAI" two years ago, that's a big change. The usual caveat applies. These are almost all Microsoft's own comparisons, mostly given as percentages of a competitor. Hacker News was cool on it. One person called the keynote "shallow," and another argued Microsoft needs open weights to win developers. But a daily MAI Code 2 user called it "quite good," roughly on par with the top models. That's a decent review from someone actually using it.

Kate

OpenAI published a post on using its own models to design chips. Marcus, which chip?

Marcus

OpenAI's page blocked us, so this comes from secondary coverage. It's tied to Jalapeño, OpenAI's first custom inference chip, built with Broadcom and unveiled earlier this year. Reports say OpenAI's models helped with the design, and that tape-out, the point where a design is sent off to be manufactured, took about nine months. That's fast for a first chip. Axios says a limited number of Jalapeño systems are running this year, more next year, and OpenAI has no plans to sell it to others.

Kate

So the AI helps build the hardware that runs the AI. That feels like a loop.

Marcus

It is, and it's the interesting part. Chip design cycles are usually measured in years, so shortening them matters a lot to anyone short on compute, which is everyone. One Hacker News commenter raised a sharper point. Chip companies are OpenAI's suppliers and potential customers, and OpenAI is keeping these design tools to itself. Labs increasingly compete with the companies they'd normally sell to.

Kate

Something more fun. Google DeepMind's Genie 4 is out through Project Genie. Marcus, remind people what a world model is.

Marcus

It's a model that generates an interactive environment frame by frame as you move through it. No game engine and no level designer. You walk around, and the model draws what you'd see. What's new in Genie 4 is that you can type prompts in the middle of a session, like making it rain, adding a character or changing the plot. It also generates real-time audio that fits the scene, from ambient sound to character dialogue.

Kate

So it's a dungeon master that also does the sound effects.

Marcus

Pretty much. It's rolling out to eligible Google AI Ultra subscribers, and Pro members can join a waitlist. There's an agent that builds worlds from a conversation, a community gallery for remixing other people's worlds, and SynthID watermarks on the output. Google is upfront about the weak spots: camera lag, complex physical interactions, and keeping a scene consistent over a long session.

Kate

How long can you play?

Marcus

Google hasn't said for Genie 4, and it hasn't given pricing either. Genie 3 sessions top out at three minutes. That was the main complaint on Hacker News. One person asked how something this compute-hungry could ever make money. It's a fair question, but it's also what people asked about video streaming in 2007.

Kate

AI agents doing an audit. KPMG UK ran something called Project Exponential. Marcus?

Marcus

They put more than a hundred AI agents on the financial statements of an unnamed FTSE 250 company, alongside the normal human audit team. It took about four months to build. KPMG UK's head of audit, Cath Burnet, said the agents sometimes probed deeper than the humans and raised questions the auditors hadn't thought of.

Kate

So should auditors be worried?

Marcus

Not yet, and the caveats matter. It was an experiment run alongside the real audit, and its findings weren't used to sign off the accounts. At one point a human had to step in because an agent misapplied a regulatory requirement. KPMG itself says it isn't ready for thousands of engagements, and the UK regulator has said the human auditor stays responsible. And Hacker News made a good point: asking more questions isn't the same as knowing which answers are material.

Kate

But they're expanding it.

Marcus

To up to twenty larger audits. The question I'd ask is whether clients get lower audit fees if this works. KPMG hasn't said, and I suspect clients will ask.

Kate

A research result. Energy-Based Reasoners, from Duke, UIUC and Harvard. Marcus, the headline says three hundred and eighty-five times less compute?

Marcus

The project page actually says three hundred and forty-five times, on a very hard Sudoku benchmark. The model has just one point eight million parameters, tiny next to a chatbot. Instead of writing out an answer word by word, it learns a score for how wrong a candidate solution is, then keeps adjusting the solution to lower that score. It stops when it stops improving. It gets ninety-nine point four percent on Sudoku-Extreme and a hundred percent on eight-by-eight N-Queens.

Kate

Should the big labs be nervous?

Marcus

No. It's four puzzle benchmarks, some baselines were copied from other papers rather than rerun, and these models are notoriously hard to train. But it's a reminder that for problems with strict rules, a much smaller and different kind of model can beat brute force. Yann LeCun has been making that argument for years, and this gives him a nice data point.

Kate

And the top Hacker News story overnight, over seven hundred and seventy points. The founder of All About Berlin, a guide for people moving to the city, wrote a post titled "AI is killing All About Berlin."

Marcus

His point is simple. People used to search, land on his carefully researched pages, and maybe use his affiliate links. Now they ask a chatbot, which gives them an answer likely built partly on his work, and they never visit. He says next year may be his last running it full time. We only have the summary, not the full post, but the thread turned into a big argument about AI training on creators' work without paying them.

Kate

That's sad, honestly. That's exactly the kind of site the internet is for.

Marcus

It is, and his problem is a business-model problem more than a legal one. If AI answers end up depending on independent experts who can no longer earn a living, the answers get worse too. The labs have a selfish reason to work out how to pay the people they rely on, through licensing, referral traffic, or something we haven't seen yet.

Kate

One to watch: Anthropic's first trade, today. Where that opening price lands compared with eighty-three dollars will tell us whether public investors believe the frontier-lab story, or whether the bankers just sold it well.

Marcus

I'd look past day one. The first few weeks, once employees and early investors are able to sell, will tell us more than the opening.

Kate

That's your AI in 15 for today. See you tomorrow.