Quick answer
Anthropic’s record IPO is this week’s biggest AI news, but the real test comes later, when insiders are allowed to sell their shares. In his weekly AI news round-up, Matt Wolfe explains why the staged share unlocks, starting in mid-December, could show whether the people closest to AI believe the boom will last. He also tests three new models (GPT-6.2 Astra, GPT-6.2 Sol and Gemini 4 Neon), finds that two of them built almost the same website, and covers Microsoft’s cheaper MAI models, Gemini Notebook’s website builder, HeyGen Avatar 6 and more.
Key takeaways
- Biggest IPO ever: Anthropic raised about $81.3 billion at $83 a share and opened on the Nasdaq (ticker ANT) at $111.50.
- Watch the unlocks: employees could sell up to 10% of their stock on day one; the next batch unlocks in mid-December, with more from early January to early February.
- The theme of the week: new models that are nearly as smart as the best for much less money.
- Neon impressed Matt most on value: close to Astra on Artificial Analysis at a lower price, and top of its coding agent index.
- AI websites are starting to look the same. Astra and Sol built near-identical deep-sea sites from an open “impress me” prompt.
In this guide
- How big was Anthropic’s IPO?
- Why are Anthropic’s share unlocks the real AI bubble test?
- What new AI models launched this week?
- Which model built the most impressive website?
- What did Microsoft announce at Ignite?
- Can Gemini Notebook really build a website?
- What else happened in AI this week?
- What should you do with this week’s AI news?
- Watch: the source video
- Frequently asked questions
- Related posts
How big was Anthropic’s IPO?
Anthropic, the company behind Claude, priced its IPO on Wednesday evening (US time) at $83 a share, selling 980 million shares and raising about $81.3 billion. That beats the previous record, which Matt says was SpaceX’s listing in June. Trading began on Thursday 19 November on the Nasdaq under the ticker ANT. It opened at $111.50 and, according to Matt, closed the first day at about $97, roughly 17% above the IPO price.
Valuation figures vary depending on the share price and share count used, but every report puts the company well above $1 trillion. We first explained why this listing matters in our earlier article on Anthropic’s IPO filing.
| IPO fact | Figure |
|---|---|
| IPO price | $83 per share |
| Shares sold | 980 million |
| Money raised | about $81.3 billion |
| First trade | $111.50 (Nasdaq: ANT) |
| First-day close (per Matt) | about $97, up roughly 17% |
Why are Anthropic’s share unlocks the real AI bubble test?
Matt’s point is that the first-day price is not the story. Normally, insiders (employees and early investors) are “locked up” and cannot sell for about six months after an IPO. Anthropic is doing it in stages instead:
- Day one: employees could sell up to 10% of their stock, up to about 315 million shares.
- Mid-December: the next batch unlocks, about four weeks after the IPO. This is the date Matt is watching first.
- 5 January to 2 February: three more unlock windows. By then, up to about 1.6 billion insider shares could be sold, more than were sold in the IPO itself.
- Within six months: most of the rest unlocks, with co-founders and the largest shareholders waiting up to a year.
These are the people who know the company best. If they sell heavily at the first chance, it could signal doubt about where AI is heading, and that doubt could spread across the AI market. If they mostly hold when they could cash out billions, the best-informed people are effectively saying it is still early.
Two other things change now Anthropic is public. It must publish results every quarter, so we will see real numbers rather than leaks. Its filing already showed about $16.3 billion of revenue in the first half of the year, up from about $1.2 billion a year earlier, and $8.9 billion spent on computing. And there is new pressure from shareholders who want growth. Matt also notes that OpenAI’s leadership has said it plans to go public, possibly in 2027.
⚠️ Warning: None of this is investment advice. Share unlocks can move a price for many reasons, and we would not read too much into any single day. If you are thinking about buying shares, speak to a regulated financial adviser.
What new AI models launched this week?
Three new models arrived this week, and they all share the same pitch: nearly the best, for less.
| Model | Tier | API price per million tokens (input / output) | Artificial Analysis score |
|---|---|---|---|
| GPT-6.2 Astra (OpenAI) | Top model | $8 / $40 | 62 |
| GPT-6.2 Sol (OpenAI) | Everyday workhorse | $2 / $10 | — |
| Gemini 4 Neon (Google) | Middle “workhorse” tier | $3 / $15 | 61 |
| Claude Fable 5.2 (for comparison) | Top model | $8 / $40 | 65 |
GPT-6.2 Astra and Sol are mostly about efficiency: fewer tokens to finish the same job. A benchmark task on Artificial Analysis cost about $4 on Astra versus about $26.54 on Fable 5.2. We cover this fully in our GPT-6.2 Astra review.
Gemini 4 Neon sits between Google’s big Carbon model and the small, fast Hydrogen, roughly where Claude Opus and GPT-6.2 Sol sit. It has a 1-million-token context window, cheap cached input ($0.15 per million), and is in the Gemini app for AI Pro and Ultra subscribers as well as the API. On Artificial Analysis it scores 61, just one point behind Astra, and Google says it costs 32% less per task than Carbon while scoring a point higher. It also tops Artificial Analysis’s coding agent index (59) when paired with Google’s Antigravity coding tool. Matt calls it “probably the best bang for your buck right now”. For full test results, see our Gemini 4 Neon review.
💡 Tip: On Matt’s SVG drawing test, Neon scored 7.1 for about 70 cents, against Astra’s 7.4 for about $1.42. On his 3D game test, Neon nailed the gameplay but Astra had better graphics. Price and quality trade off, so test on your own tasks.
Not sure which model to pay for? Our guide to testing whether a cheaper AI model is good enough helps you compare fairly on your real work.
Which model built the most impressive website?
Matt has brought back an old test: give each model the same open prompt asking it to build a website that will “truly impress me”, on any subject it likes. The idea is to see which models have taste when nobody is steering the design, because, as he says, AI-coded websites are starting to look the same.
- Astra built a deep-sea “ocean atlas” with jellyfish, depth zones and a “begin descent” button. Nice, but not very interactive.
- Sol built… almost the same thing: another deep-sea site with depth zones, a “begin the descent” button and similar wording.
- Neon went somewhere completely different: a luxury watchmaker’s site with a mechanical watch you can view assembled, as a skeleton or exploded, plus a magnifying “loupe” hover effect. Matt thought it was the most impressive of the three.
Matt links this to a complaint from developer Theo Brown that Astra “regresses to the mean”: he and his podcast co-host asked it to design game levels in different games and both got a water, music and bell-themed level with the same name. When models fall back on the same safe ideas, your site can end up looking like everyone else’s.
💡 Tip: Want an AI-built site that looks like your brand? Do not leave the design to chance. Our guide shows how to stop an AI-built website looking like every other one.
What did Microsoft announce at Ignite?
Microsoft held its Ignite conference. Most of it was for big companies, but a few releases are relevant to smaller builders. All figures are Microsoft’s own:
- MAI Code 2, Microsoft’s own coding model, which it says delivers 98% of GPT-6.1 Sol’s performance at 40% of the cost, and 96% of Claude Opus 5.5 at 5% of the cost. It is aimed at simpler coding jobs rather than long, complex projects. It is in GitHub Copilot now, with OpenRouter and Vercel “coming soon”.
- MAI Image 2.7, which tops Artificial Analysis for image editing and sits fourth for generating images from scratch.
- MAI Realtime 1, a real-time voice model from about one cent per audio minute.
Can Gemini Notebook really build a website?
Yes, in a simple way. Gemini Notebook (the tool formerly called NotebookLM) can now turn your uploaded notes, documents and research into an interactive website: type “make me a site” and describe what you want. Google’s examples include sales dashboards, task trackers and study sites. Matt loaded the week’s news into a notebook and, about two minutes later, had a newspaper-style site with charts, tables and filterable sources. Some news sites blocked it from reading their pages, so not everything got in. For a step-by-step walkthrough, see how to turn your notes into an interactive website with Gemini Notebook.
⚠️ Warning: Check any AI-generated summary site before sharing it. Matt’s auto-generated headline quoted a different valuation from the figures he gave in the video, which is a good reminder that AI can mix up numbers.
What else happened in AI this week?
| Update | What it does | Who can use it |
|---|---|---|
| HeyGen Avatar 6 | Talking video of you from one photo, no set length limit, up to 2K. In HeyGen’s own test, viewers picked the AI clip as the real one 42% of the time | Released this week (not yet tested by Matt) |
| Suno Visuals | Lyric videos, album art and clips (up to 15 seconds) that follow your song | Web, iPhone and Android; separate credits |
| Gemini on Mac computer use | Clicks, types and fills in forms across your apps in a background window; asks before big actions | US only, Apple Silicon Macs, paid Google AI plans |
| Claude Duet | Anthropic’s always-on assistant opened to US waitlist users on Pro, Max and Ultra | Counts against your normal Claude limits |
| Claude Code custom themes | Describe a look and Claude rebuilds its own desktop interface | Local desktop sessions only |
| Google Stitch Loop | Keeps checking your app’s code for security, speed and accessibility problems and proposes fixes for you to approve | Waitlist; price unknown |
| ChatGPT Wallet | Lets ChatGPT’s assistant buy things using a stand-in card number, with Face ID or fingerprint approval | US only, credit cards, Pro plan |
| ChatGPT Live Dictation | Your words appear as text while you speak | ChatGPT users |
| ChatGPT for Sales | Ready-made sales workflows connecting to 30+ tools such as HubSpot and Salesforce | Business and Enterprise workspaces |
| Genie 4 (Google DeepMind) | Explorable worlds you can change with new prompts, with generated sound; Genie 3 worlds extended to 3 minutes | Google AI Ultra; Pro waitlist |
Stitch Loop tackles a real problem: lots of people now “vibe code” apps without thinking about security. You do not need to wait for it. Our guide shows how to check a vibe-coded app for security problems before launch. For more on AI agents spending money, see how to let an AI agent make purchases safely.
Sponsor note: part of the video is sponsored by HyperAgent, a platform for building teams of AI agents, which has added iMessage support and a free plan. We mention it for transparency, not as a recommendation.
What should you do with this week’s AI news?
- Do not panic or cheer over one share price. The interesting dates are the unlocks from mid-December onwards.
- Re-check what you pay for AI. Cheaper “nearly the best” models like Neon and Sol may handle most of your everyday work.
- Give AI a design brief. If two models build the same site from the same prompt, yours will look generic unless you steer it.
- Check what you build. Quick security checks on AI-built apps are now essential, not optional.
- Try one new tool. Gemini Notebook’s site builder is the easiest win this week for non-coders.
🛠️ Practical guides
Ready to put this into practice? These step-by-step guides answer two common follow-up questions:
- How to Stop an AI-Built Website Looking Like Every Other One — a simple design brief that stops AI tools falling back on the same generic layouts.
- How to Check a Vibe-Coded App for Security Problems Before Launch — a beginner-friendly checklist to catch the most common security gaps in AI-built apps.
Watch: the source video
What this video covers: Matt Wolfe breaks down Anthropic’s record IPO and staged share unlocks, tests GPT-6.2 Astra, GPT-6.2 Sol and Gemini 4 Neon (including an “impress me” website test), and rounds up Microsoft’s MAI models, Gemini Notebook sites, HeyGen Avatar 6, ChatGPT Wallet, Genie 4 and more.
Video: Matt Wolfe (@mreflow), “AI News: This Is The Make or Break Week!!”, published 20 November 2026. The HyperAgent segment is sponsored. This summary and the opinions in it are our own; tests, predictions and opinions in the video are Matt’s, and performance figures are the companies’ own claims unless stated.
Next step: pick one website or tool you are building with AI and spend 15 minutes on our pre-launch security check. It is the cheapest insurance you will buy this week.
Frequently asked questions
How much did Anthropic raise in its IPO?
About $81.3 billion, selling 980 million shares at $83 each, the largest IPO on record according to Matt Wolfe. Shares began trading on the Nasdaq under the ticker ANT on 19 November 2026.
What is an IPO lock-up or share unlock?
After an IPO, insiders such as employees and early investors usually cannot sell for a set period. Anthropic is releasing them in stages: some on day one, more in mid-December, and further windows from January.
Is Gemini 4 Neon better than GPT-6.2 Astra?
Not on raw intelligence: it scores 61 on Artificial Analysis against Astra’s 62. But it is cheaper, tops the site’s coding agent index and impressed Matt in his website test, so it may be better value.
Why do AI-built websites look the same?
Models often fall back on safe, popular ideas when the prompt is vague. In Matt’s test, Astra and Sol built almost identical deep-sea sites. A clear brand and design brief fixes most of this.
Is there an AI news video next week?
Matt says there will be no news video in Thanksgiving week while he is on holiday, though he will cover anything big when he returns.
Related posts
- GPT-6.2 Astra Review: Cheaper and Efficient, but Is It Worth Switching?
- Gemini 4 Neon Review: What the Tests Show on Speed, Cost and Quality
- Anthropic IPO: Why It’s an AI Bubble Test, Plus Kimi K3.1 Pro, Genie 4 and OpenAI’s Maths Claim
- How to Turn Your Notes Into an Interactive Website With Gemini Notebook
