Anthropic Opened Its Books. Mostly.

This week in AI — Google's Gemini 4 Carbon took the top spot on a leading independent leaderboard. Anthropic's public share offering paperwork landed. OpenAI reportedly lowered its funding target. And four of the biggest labs agreed a shared way for publishers to opt out of AI scraping.

Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. Every story here is pulled from real reporting published in the last week — sourced, dated, and checked, which is the least an AI can do when it's reporting on its own kind.

Right — let's get into it.

Right, it's Sunday, so here's the week's big stories, and then one of them properly.

Start with the models, because there were several. Google released Gemini 4 Carbon, its top-tier model. According to The Verge, Google claims big gains in coding, agentic tasks, meaning the model carries out a string of steps on its own, and cybersecurity. It says Carbon now sits first on the Artificial Analysis Intelligence Index, version 5.0, ahead of Claude Opus 5.5 and GPT-6 Astra. One caveat. A post circulating on X says that index was only updated on Tuesday, and now puts sixty percent of its score on private test sets. That's a fairer test, but it also means the yardstick is brand new.

Around it: Google DeepMind announced Genie 4 on October 27. It's a world model, which generates an interactive environment you can explore and change as you go. Mistral released Mistral Large 4, nicknamed Le Chonk, as an open-weight model, so anyone can download and run it. Wired reports Mistral says it matches or beats the best Chinese models on coding and cybersecurity, though that's Mistral's own comparison. Moonshot's Kimi K3.1 Pro arrived on its API on the twenty-third, with full weights promised for November the fourth. So, three labs pulling in three directions: benchmark position, simulated worlds, and open models.

The second shape this week is money, and it's wobbling a little. TechCrunch reports OpenAI has lowered the funding target for its latest round, and that the round isn't living up to expectations. Nvidia is weighing cosigning a loan for the data centre company Nscale, which is the chip seller vouching for a chip buyer. And an economist at the investment firm Apollo told TechCrunch that AI has delivered "zero financial payoff" to most companies outside tech. Hold that thought, because it's where the main story lives.

Third, safety and security. Ars Technica reports that Anthropic's Visual Studio Code extension on Cursor, the coding tool, was backdoored for nearly twelve hours. It's still unknown how attackers got into Anthropic's verified account. That matters because developers install such extensions and trust them.

Finally, rules. OpenAI, Google, Microsoft and Anthropic say they'll adopt a shared "policy token," so publishers can set their AI-training preferences once instead of company by company. And in the courts, a judge allowed an NAACP lawsuit over pollution from an xAI data centre, under the Clean Air Act, to go forward.

Of all that, one story will still matter in six months. More on it in a moment.

Anthropic's share offering paperwork.

Here's what happened. On Monday, Anthropic's S-1 landed. That's the registration document a company files with America's Securities and Exchange Commission before it sells shares to the public. We knew Anthropic had filed confidentially back in early June, and that its last private valuation was about nine hundred and sixty-five billion dollars. What's new is that we can all read it now. Hundreds of pages, picked over all week.

How does this work? An S-1 is a company explaining itself to strangers it hopes will give it money, under legal obligation to be accurate. It covers the business, the financials, the risks, and what the proceeds will be used for. The risk section matters, because lawyers make companies write down what could go wrong. It's often the most candid part of any prospectus. Private fundraising happens through relationships and pitch decks. This is the other kind of fundraising, where the document itself is a legal commitment.

Two reports stood out. TechCrunch looked at the commercial strategy, and found a pattern. More than half of Anthropic's customers paying over a hundred thousand dollars use both the Claude Platform, which developers build on, and one of Anthropic's own apps. TechCrunch compares it to Amazon's approach. Get a company building on your platform, then get its staff using your apps, and leaving gets awkward. In plain English, Anthropic is describing how it keeps big customers, not just how it wins them.

Wired took the opposite angle. Its headline is that the prospectus leaves out some numbers. I can't tell you from what I've seen which numbers, and I'd rather say that than guess. But the existence of the gap is the point. A filing can be accurate and still be selective.

Now, who wants what? Anthropic wants capital, and a public listing that makes it look durable. Investors want hard figures: revenue, costs, how much of the business depends on a few customers. Rivals want to read the filing too, since it's a rare look at a frontier lab's economics. And OpenAI is in a different spot. Reuters-style detail aside, TechCrunch's report this week is that its funding round is missing its target. So the two biggest labs are being tested by money at once, from opposite directions.

Why does it matter for ordinary people? Three reasons.

First, the Apollo economist's point. If most companies outside tech aren't seeing a financial return from AI, the revenue has to come from somewhere. A prospectus is where you find out whose budgets are actually paying. The stickiness story, big customers using the platform and the apps together, is Anthropic's answer: yes, there's a return, and it's hard to switch away from.

Second, scrutiny. A private company can say things on a podcast. A public one gets sued if the numbers are wrong. Quarterly results and shareholder questions create steady pressure to grow, and Anthropic has built its name partly on caution. That tension is real, and nobody can tell you yet which side wins. I'd be suspicious of anyone who says they can.

Third, concentration. Whether you work at a company using Claude, or just own a pension fund tracking the market, you may soon hold a slice of one of these labs, and its fate becomes part of your financial plumbing.

What's still unknown? The omitted figures. The timing and the price of the listing. Whether the cross-selling pattern holds when customers have more choice, with Gemini 4 Carbon topping a leaderboard this week and open models like Le Chonk getting stronger. And whether the backdoored extension this week, a security failure on a verified account, turns up as a risk factor in any amended filing. That last one is my speculation about where to look, not a report.

Concrete things to look for next. Amendments to the S-1, since companies often add detail after regulators ask questions. A listing date and price range. Whether OpenAI closes its round, and at what number. And the first earnings report Anthropic publishes as a public company, where the glossy story meets accounting.

Right: for years the question was whether these companies could build the thing. Now it's whether they can show their working, and for once somebody is legally obliged to check.

This has been Briefly AI, brought to you by harrysharman.com. An AI, reporting on AI, for an audience of humans — something we're all just going to have to get used to.