Anthropic IPO: Why It’s an AI Bubble Test, Plus Kimi K3.1 Pro, Genie 4 and OpenAI’s Maths Claim

What is happening with the Anthropic IPO?

On Monday Anthropic, the company behind Claude, filed its S-1 with the US Securities and Exchange Commission. An S-1 is the document a company must publish before it can sell shares on a stock exchange. It is often the first time outsiders see real figures for revenue, spending and who owns what.

According to Matt Wolfe’s video, the listing is expected around 16 November at a valuation of somewhere between $1.5 trillion and $2 trillion. The filing itself runs to roughly 185 pages. We looked at what it means for business buyers of AI in our earlier piece on what the Anthropic S-1 reveals about enterprise AI adoption; this article focuses on the money and the bubble question.

What do Anthropic’s numbers actually show?

In half a year, Anthropic earned almost four times what it made in the whole of 2025. That is remarkable growth for a company of this size.

The loss needs context. Matt explains that some early deals involved investors lending Anthropic money that will later turn into shares. Under accounting rules, changes in the value of that debt can show up as a loss even though the cash never left the bank. So the headline loss is not the same as money spent, although Anthropic does still spend heavily on computing power.

Who gets rich if Anthropic lists at $2 trillion?

Matt shared a breakdown posted on X by Andrew Benson. At a $2 trillion valuation, it estimates roughly $118 a share, with:

• about $199 billion going to the seven co-founders,
• about $100 billion to Anthropic’s public benefit fund,
• about $331 billion to staff through share options and restricted stock, an average of around $62 million for each of roughly 5,300 employees.

The same post estimates that around 940 people could end up with nine-figure sums, with the largest amounts going to those hired by 2023. Early secondary-market buyers reportedly paid $25.91 a share in April, and a later funding round was priced at $58.90.

Large cloud and chip partners such as Amazon, Google and Broadcom also stand to gain, because they supply computing power to Anthropic, sometimes partly in exchange for shares.

Why is the Anthropic IPO a test of the AI bubble?

Google, Microsoft and Meta all sell AI, but AI is only one part of each business. Anthropic would be the first company that is almost entirely an AI lab to trade publicly, so its share price becomes a running scorecard for confidence in AI.

Matt’s point is about timing. Insiders usually cannot sell on day one. A lock-up period, often around six months, keeps them holding. When it ends, we will learn something real:

• If early staff and investors hold on, it suggests the people closest to the technology think it has much further to grow.
• If they sell heavily, it suggests many were in it for a quick payday rather than a long-term belief in AI.

Who is GPT-6.1 Terra actually for?

OpenAI now has four model families. In Matt’s summary, Astra is the top tier, then Sol, then Terra, then Luna as the smallest and cheapest. He compares Sol to Anthropic’s Opus, Terra to Sonnet and Luna to Haiku. GPT-6.1 Terra is the newest of these, sold as faster computer use at a lower price.

The trouble is that, on the charts Matt walked through, Terra rarely wins on value:

Terra is a little cheaper per task and only very slightly faster. Matt’s view is that if you want cheap and smart you would pick Sol, if you want the best you would pick Astra, and if you want tiny and fast you would pick Luna. We covered the launch itself in our GPT-6.1 Terra round-up.

Is Gemini 4 Carbon now the top model?

On paper, yes, but you probably cannot use it yet. Google has announced Gemini 4 Carbon, a step up from Gemini 4 Argon, and it now sits at the top of the Artificial Analysis leaderboard, ahead of Claude Opus 5.5 and GPT-6 Astra.

For now it is only open to early testers and some security researchers, with wider access for developers, businesses and consumers “coming soon”. Matt notes it looks cheaper than most of Anthropic’s models but more expensive than most of OpenAI’s. Our Gemini 4 Carbon article has more detail on early tests.

Can you run Kimi K3.1 Pro at home?

Realistically, no. Kimi K3.1 Pro, from the Chinese lab Moonshot AI, is set to become the strongest open-weight model when its weights are released on 4 November. Open-weight means you can download the model and run or fine-tune it yourself, rather than only using it through the company’s app.

• 2.8 trillion total parameters, with about 104 billion active for each answer.
• A one-million-token context window.
• An Artificial Analysis score of 46, only about three points behind Claude Fable 5.1, which caused alarm just a few months ago. The next-best open models sit in the low 30s.

The catch is size. All those parameters must sit in memory. Matt estimates you would need around 44 Nvidia RTX 5090 graphics cards (32 GB each), or at least three fully loaded Mac Studios with 512 GB of memory, at roughly $10,000 to $15,000 each.

So the privacy benefit of open weights is real mainly for organisations such as hospitals, banks and governments that can run their own servers. For everyone else, the value is that smaller, more efficient open models keep improving. If you want to know what you could run, our new guide explains how much memory you need to run an open-weight AI model.

What is Genie 4 like to use?

Genie 4 from Google DeepMind is a “world model”: you describe a scene and a character, then move around it in real time with your keyboard. Matt got early access (there is a waiting list) and tried a rainy neon city and a cartoon sweet-shop world.

• New this time: you can prompt changes while you play, such as a police drone flying in or a giant monster crashing through a wall, and the worlds now have sound and music.
• Limits: sessions stop after about two minutes, one session froze, and scenery can change when you turn around and look back.

Matt does not see it as a game yet. He sees likely uses in filmmakers planning shots, game designers testing ideas before building them properly, and robotics companies training robots in simulated worlds. We first covered Genie 4 in our Genie 4 news round-up.

Did OpenAI’s AI really disprove a 25-year-old maths assumption?

That is OpenAI’s claim. It says an unreleased internal model found a disproof of the Exponential Time Hypothesis on 16 October. This is a roughly 25-year-old assumption in computer science that the hardest versions of certain puzzles always take exponentially longer to solve as they grow.

Matt’s simple picture: imagine a row of light switches and a long list of rules such as “if switch three is on, switch seven must be off”. Your job is to find a setting that breaks no rules. Every switch you add doubles the number of combinations, so with a few hundred switches there are more possibilities than atoms in the universe. Computers face puzzles like this in scheduling, chip design and software testing.

Hundreds of research papers assume the hypothesis is true. OpenAI says its proof was checked in Lean, software that verifies every step of a proof follows from the one before, like a very strict spell checker for maths.

Why it matters: most AI so far has been good at things people already know how to do. This would be AI answering a question experts had not cracked. Matt suggests reading OpenAI’s article and asking an AI assistant to explain it back, which is a good habit. Our guide on using AI to explain a complex topic in plain English shows how to do that without being misled.

What should you do with this news?

1. Ignore the IPO noise for day-to-day decisions. Your tools matter more than share prices, but expect suppliers to keep adjusting prices and limits.
2. Keep your workflows portable. Save prompts, instructions and examples in your own documents so you can switch models.
3. Test before you switch. New models such as Terra are not automatically better value.
4. Watch open-weight models, but be realistic about the hardware they need.
5. Read big claims carefully. Separate what a company says from what has been independently checked.

Matt also says his news videos will now focus on the handful of stories that really matter, with the full list of weekly updates on his Future Tools news page. That is a sensible way to follow AI without drowning in announcements.

Frequently asked questions

The bottom line

The Anthropic IPO is the headline, and it really will be a test of how much people believe in AI. For your business, the practical lessons are quieter: test new models on your own work, keep your workflows portable, be realistic about open-weight hardware and read big claims with care. Your next step: pick one AI task you run every week and note which model you use and what it costs, so you are ready when prices change.

Related posts

• Anthropic S-1: What It Reveals About Enterprise AI Adoption
• GPT-6.1 Terra: OpenAI’s Cheaper Fast Model, Anthropic’s IPO Filing and This Week’s AI News
• Kimi K3.1 Pro Review: Is It Really Cheaper Than Claude Opus 5.5?

Garry Knight

I'm Garry Knight, the person behind Prodify Digital. I write about email list building, email marketing, SEO, AI search and the tools that connect them. My aim is to make online marketing easier to understand, so creators and small business owners can make informed decisions about building an audience and keeping people engaged. Here you'll find straightforward guides and product reviews that explain what something does, where it fits and which limitations matter. The focus is on clear explanations and useful next steps—not hype, shortcuts or promises of easy earnings.

See Full Bio