The AI Spotlight Follow
Anthropic Just Pulled Off the Biggest IPO Ever. The Real AI Bubble Test Starts on December 15
AI Money

Anthropic Just Pulled Off the Biggest IPO Ever. The Real AI Bubble Test Starts on December 15

Anthropic raised $81.34 billion and closed its first day up 17%. But the number that matters most isn't the share price. It's a 23-step insider unlock calendar that begins in December. Here's how to read it, plus the rest of a packed week: GPT-6.2, Gemini 4 Neon, Microsoft's cheap MAI models, ChatGPT Wallet and more.

Advertisement

On Wednesday night, November 18, the company behind Claude sold a slice of itself to the public for $81.34 billion.

That is the largest initial public offering in history, larger than SpaceX's record-setting debut earlier this year. On Thursday the stock opened on the Nasdaq under the ticker ANT at $111.50, slid through the afternoon and still closed at $97.17, about 17% above the offering price.

The headlines wrote themselves. But the first-day pop is not the interesting part.

The interesting part is buried in the prospectus: a staged schedule that decides when the people who know Anthropic best, its employees and early investors, are allowed to sell. Over the next year, up to 15.5 billion locked shares become eligible in 23 separate steps. The first big one after listing day lands on December 15.

Advertisement

What those insiders do with that freedom may tell us more about whether AI is a bubble than any analyst note. Here is how the deal worked, what to watch and when, and the rest of a week that also brought three new frontier models, a wave of cheap Microsoft AI and a wallet that lets ChatGPT spend your money.

The deal, in numbers

Here are the verified facts from Anthropic's pricing announcement and market coverage:

  • Price: $83 per share, above the $70-$78 range Anthropic had marketed.
  • Shares sold: 980 million Class A shares. Anthropic itself sold about 956 million; the co-founders sold about 23.6 million between them.
  • Gross proceeds: $81.34 billion.
  • Demand: Morningstar reports orders came in at more than six times the 825 million shares originally planned, which is why the deal was upsized.
  • Extra: underwriters have a 30-day option on another 147 million shares.
  • First day: opened at $111.50 just before noon ET, closed at $97.17.

We covered the roadshow and the original price range when it launched in AI's wildest week. This is the payoff, and it came in hotter than the range suggested.

Why the first-day price isn't the story

First-day pops are mostly about allocation and hype. Banks price a deal to leave some money on the table, retail buyers pile in at the open, and the price settles. It tells you about demand on one day.

The better signal arrives later, when insiders are free to sell.

In a traditional IPO, almost everyone who held stock before the listing is locked up for about 180 days. Then a single, well-known date arrives, a big block of supply becomes available at once, and markets brace.

Anthropic did it differently. Its lock-up releases shares in stages, and it started on day one.

Anthropic's unlock calendar

According to the final prospectus, as tabulated by analysts at Perpys and CrowdfundedWealth, the schedule looks like this:

When Up to (shares) What it is
Nov 19, 2026 (first trading day) 315 million Non-founder directors, officers, employees and consultants could sell up to 10% of their locked stock
Dec 15, 2026 303 million First calendar release after listing
Jan 5, 2027 303 million
Jan 19, 2027 353 million
Feb 2, 2027 353 million Last of the five fixed-date releases
After FY2026 results 1.7 billion, then five steps of 480 million Tied to the first annual report
After Q1 2027 results or May 17, 2027 3.0 billion End of the 180-day group; the largest single release
Through Nov 18, 2027 Several more steps Co-founders and holders above 0.5% wait up to a year

The five fixed dates alone add up to about 1.6 billion shares, well over the 980 million sold in the IPO itself.

Two cautions matter here:

  1. These are ceilings, not predictions. "Eligible to sell" doesn't mean anyone will sell.
  2. Some selling is normal. Employees buy houses, pay tax bills and diversify. Venture funds have to return money to their own investors. Modest selling is not a vote of no confidence.

What matters is the size and speed of selling relative to what's allowed.

How to read the unlocks

Think about who holds these shares. They are the engineers who built Claude, the executives who see the internal metrics and the investors who have read every board deck for years. Nobody has better information about where Anthropic, and frontier AI more broadly, is heading.

So the coming unlocks create a natural experiment:

  • If insiders mostly hold when they could walk away with life-changing money, the people with the most information are saying they think we're still early.
  • If insiders and early funds sell heavily at the first chance they get, that's a quiet signal that people close to the technology think prices have run ahead of reality. That kind of doubt can spread quickly across AI stocks.

You won't see individual employee trades in real time. But you will see the stock's behavior around each release date, trading volume, Form 4 filings for officers and directors, and any public commentary from big funds. December 15 is the first clean data point.

Two more things change now that Anthropic is public

1. Real numbers, every quarter. As a private company, Anthropic's revenue and spending reached the public through leaks and selective disclosures. Now it must publish audited results on a fixed schedule. That will finally show how fast revenue is growing relative to the colossal cost of compute, which is the central question behind every "AI bubble" argument.

2. A stock price that never sleeps. Shareholders want growth. That pressure will shape release timing, pricing and how the company talks about safety. It's one reason critics paid close attention when, a day after the debut, an Anthropic mathematician announced a claimed proof of the Riemann Hypothesis made with Claude. We break down that claim in Claude and the Riemann Hypothesis.

And the company watching most closely is OpenAI. Its CFO has reportedly told staff it plans to go public in 2027 or sooner, and Anthropic's reception, plus how its insiders behave, effectively sets the template.

The week's three new models

Anthropic's listing overshadowed what would normally be the week's lead: three frontier-class releases.

GPT-6.2 Astra and Sol

OpenAI shipped GPT-6.2 Astra, its top model for the hardest work, and GPT-6.2 Sol, the cheaper everyday model. Both are in ChatGPT and Codex for paid plans and in the API.

The pitch is less "much smarter" and more "same work, fewer tokens." On coding benchmarks, 6.2 Astra edges past 6 Astra, while 6.2 Sol lands close to 6.2 Astra at a much lower price. Independent rankings put Astra second overall behind Claude Fable 5.2, but at a fraction of Fable's cost per task.

Hands-on testers found Astra strong on visual work: polished 3D scenes and game prototypes. The complaints are consistent, too. Early users say it can stop too soon and be hard to follow on big projects, and some developers say it "regresses to the mean," giving different people strikingly similar outputs. In one informal test where Astra and Sol were each asked to "build a website that will truly impress me," both produced near-identical deep-sea themed sites, down to similar wording.

If you want the deep dive, read GPT-6.2 Astra tested.

Gemini 4 Neon

Google's Gemini 4 Neon slots between Carbon (the flagship) and Hydrogen (the small, fast model) at $3 per million input tokens and $15 per million output, with a million-token context window. It's in the Gemini app for paid subscribers and in the API.

What makes it notable is value. Independent benchmarking puts it within a point or so of GPT-6.2 Astra at well under half the price, and Google says it costs 32% less per task than Carbon. On an agentic coding index, the Neon-plus-Antigravity combination ranked first at launch. In the same "impress me" test, Neon went somewhere different entirely, a mechanical-watch site with an exploded view and a magnifier effect, which many viewers found more original than the OpenAI pair.

We've covered Neon twice already: the launch and pricing and a hands-on test.

Microsoft's MAI Code 2

At Ignite on November 17, Microsoft introduced MAI Code 2 and MAI Code 2 Flash, its own coding models. The claims, all Microsoft's:

  • 98% of GPT-6.1 Sol's performance at 40% of the cost, averaged across four coding benchmarks.
  • 96% of Claude Opus 5.5's performance at 5% of the cost.
  • Developers preferred its front-end results over GPT-6 Sol 64% of the time in blind tests.

Microsoft is explicit that the models are not yet tuned for hard science or very long agentic work. They're built for everyday coding: multi-file changes, debugging and UI. MAI Code 2 is available in GitHub Copilot now.

The rest of Microsoft's "almost-best, much cheaper" push

Ignite's theme was consistent: close to the frontier at a fraction of the price.

  • MAI Image 2.7 took the top spot on Artificial Analysis for image editing, and Microsoft says it beats GPT Image 2.6 Sunburst (max) at one-fifth of the price. It ranks fourth for text-to-image. Pricing is about $35 per 1,000 images, or about $10 for the Flash version, in Microsoft Foundry preview.
  • MAI Realtime 1 is Microsoft's first live speech-to-speech model, starting at $0.01 per audio minute, which Microsoft says outperforms GPT-Realtime 2.1 on its evaluations.

For small businesses, this matters more than leaderboard bragging rights. Price wars at the "good enough" tier make AI features cheaper to build into everyday products.

ChatGPT can now hold your credit card

OpenAI introduced ChatGPT Wallet on November 13, built with American Express, Mastercard and Visa, with Chase cards also supported.

How it works:

  1. You add an eligible US credit card once.
  2. Your always-on ChatGPT agent (OpenAI calls them "dots") finds what you want.
  3. Before each purchase, you review the items and total and approve with Face ID or a fingerprint.
  4. The payment network issues a payment token, a stand-in card number, so neither the agent nor the merchant sees your real card details.

It's US-only, Pro plan only (the $100-a-month tier and up), and requires a dot. OpenAI says it eventually expects users to delegate purchases within limits such as spending caps and categories.

The practical advice: if you try it, start with small, low-risk purchases, keep per-purchase approval on and watch your statements. "Agent buys things for you" is convenient right up until it misreads what you wanted.

Rapid fire: everything else that shipped

  • Gemini Notebook builds websites. The tool formerly known as NotebookLM can now turn your sources into a shareable interactive site, such as a sales dashboard, a deadline tracker or an exam study guide, from a "make me a site" prompt. In one early test, a notebook full of the week's news became a newspaper-style page with charts, tables and filterable sources in about two minutes.
  • HeyGen Avatar 6. One photo produces a talking video with no set length limit, at 2K, with examples longer than 25 minutes. HeyGen says viewers picked the AI clip as real 42% of the time in side-by-side tests. That's close to, but not yet at, the 50% coin-flip mark that would mean people truly can't tell.
  • Suno Visuals. Drag a song into the new tab to get lyric videos, album art and short clips that follow the song's structure and can lip-sync. Clips max out at 15 seconds, and visual credits are sold separately.
  • Gemini on Mac can use your computer. Describe a task and it clicks, types and fills in forms in a background window while you keep using your machine. You choose which apps it can touch, and it asks before big actions. US only, Apple Silicon, paid plans.
  • Claude Duet opens up. Anthropic's always-on assistant is now available to US Pro, Max and Ultra users who were on the waitlist. Note that it runs in the background and counts against your normal usage limits.
  • Claude Code desktop themes. Describe a look and Claude Code rebuilds its own interface to match. Fun, not productive, and local sessions only.
  • Google Stitch Loop. A waitlist tool that connects to your app and its data, keeps running security, speed and accessibility checks, and proposes fixes for you to approve. Aimed squarely at vibe coders who never think about security.
  • ChatGPT live dictation shows your words as you speak, and ChatGPT for Sales adds ready-made workflows for prospecting, call prep, follow-ups and at-risk deals across 30-plus tools like HubSpot and Salesforce, on Business and Enterprise plans.
  • Genie 4 is in Ultra subscribers' hands (we covered it in Genie 4: worlds you can hear), and Google extended its older Genie 3 sessions from one minute to three minutes per world.

What this week means for you

If you invest: Don't anchor on the first-day pop. Put December 15, January 5, January 19 and February 2 in your calendar and watch price and volume around them. Then watch the first quarterly report for the growth-versus-compute-cost picture. And remember that a single stock is not the whole AI market.

If you build: The "near-frontier at a fraction of the price" tier is getting crowded fast: Gemini 4 Neon, GPT-6.2 Sol, MAI Code 2. Re-run your cost math. Many teams can move routine work to cheaper models and save the expensive ones for hard problems.

If you're a regular user: Agents are being handed real-world powers this month: your computer (Gemini on Mac), your credit card (ChatGPT Wallet) and your background attention (Claude Duet). Each comes with approval steps. Use them.

Counterpoints

  • Unlock data is noisy. Prices move for many reasons, and a drop around a release date doesn't prove insiders sold.
  • One company isn't "AI." Anthropic's stock could struggle while AI adoption keeps growing, or soar while the broader sector wobbles.
  • Vendor benchmarks are vendor benchmarks. Microsoft's 98% and 96% figures, HeyGen's realism numbers and lab leaderboards all need independent confirmation on your own tasks.

FAQ

How much did Anthropic raise in its IPO? $81.34 billion gross, selling 980 million shares at $83, the largest IPO on record.

When can Anthropic insiders sell? In stages. Up to 315 million shares on the first trading day, then fixed releases on December 15, January 5, January 19 and February 2, then larger releases tied to earnings reports, with the 180-day group free by May 17, 2027 at the latest and co-founders by November 18, 2027 at the latest.

What is the ANT ticker? Anthropic's Class A shares trade on the Nasdaq Global Select Market under ANT.

Is ChatGPT Wallet safe? It uses tokenized card numbers and requires biometric approval for every purchase, but it's new, Pro-only and US-only. Start small.

Which new model is the best value right now? For many API users, Gemini 4 Neon offers near-top intelligence at mid-tier prices, but test on your own workload first.

The bottom line

Anthropic's record IPO answered one question, whether public investors want a piece of a frontier AI lab, with a resounding yes.

The harder question, whether the people inside believe the valuation, gets answered slowly, one unlock at a time, starting December 15. Meanwhile, the model race keeps compressing prices, and agents are quietly gaining access to your computer and your wallet. Watch the calendar, re-check your AI costs and keep your approval buttons switched on.

Sources

There's No Single 'Best AI Model' Anymore. OpenAI's Own Data Shows Why, and Why the Smartest Ones Are Getting Harder to WatchUp next · 11 min readThere's No Single 'Best AI Model' Anymore. OpenAI's Own Data Shows Why, and Why the Smartest Ones Are Getting Harder to WatchGPT-6.2 Astra, Claude Fable 5.2 and Gemini 4 Neon each 'win' depending on which leaderboard you open. The reason isn't raw intelligence; it's training data. And two new OpenAI reports add a twist: AI is now tripling some researchers' output, while its reasoning is becoming harder for humans to monitor. Here's how to pick a model for your work, and what the fine print says about where this is heading.Keep reading →