Agents, Anthropic, and a Very Awkward Book

Welcome to Briefly AI. Today: Agents, Anthropic, and a Very Awkward Book.

Anthropic just became the most valuable AI startup on the planet. GitHub Copilot just became a lot more expensive. And a book about AI hallucinations turned out to contain AI hallucinations. You cannot make this stuff up. Well, apparently someone tried. Let's get into it.

Right, let's start with the big number. Anthropic — the company behind Claude — has just closed a sixty-five billion dollar funding round, valuing the company at nine hundred and sixty-five billion dollars. Nearly a trillion. That puts them ahead of OpenAI as the most valuable private AI company in the world, which is a sentence I genuinely didn't expect to be saying this year.

Now, we covered Anthropic's thirty billion dollar round a couple of weeks ago, so this is a new development — and a significant one. That round closed. Then they immediately raised again, more than doubled it, and are now in the territory where the next logical step is an IPO. Their filings are already being described as pre-public positioning.

Here's the bit that matters though — beyond the headline number. The gap between the top-funded labs and everyone else is now structurally enormous. Sixty-five billion dollars buys an almost incomprehensible amount of compute, talent, and time. If you're a smaller lab, a startup, or a business trying to build on top of AI models, the landscape you're operating in is increasingly shaped by a handful of entities with essentially unlimited capital. That's not necessarily bad — it means faster research, more capable models, prices that come down over time — but it does concentrate an enormous amount of power in very few hands. Worth keeping an eye on, especially as both Anthropic and OpenAI start eyeing the public markets.

On a completely different note — and this one's a bit more practical, especially if you're a developer or work with developers — GitHub Copilot has just made a change to its billing model, and the reaction from the developer community has been, let's say, pointed. The phrase "what a joke" was used, in a headline, which is fairly rare.

Here's what happened. GitHub — which is owned by Microsoft — has shifted Copilot's pricing from a flat monthly subscription to token-based billing. Tokens, if you're not familiar, are the units AI models use to measure how much text they process — roughly, every word you write or read costs tokens. What this means in practice is that heavy users of Copilot — the people who use it all day, every day, for serious coding work — are going to see their bills go up. Potentially significantly.

Now, from Microsoft's perspective, this probably makes sense as a business model. Power users consume vastly more compute than casual ones, and flat subscriptions subsidise the heavy usage at the expense of margins. Fair enough. But the timing is awkward. Microsoft has just spent the last few months quietly nudging its own employees away from Claude Code and toward Copilot — which we covered recently — and now it's making Copilot more expensive for exactly the developers it needs to win over.

The deeper issue is what this signals about the economics of AI tools more broadly. The "flat fee, use it as much as you want" era might be ending. As usage goes up and compute costs remain real, providers are going to look for ways to charge proportionally. If you're budgeting for AI tools at work — and statistically someone on your team is — it's worth checking whether your costs are about to shift. What to watch: whether Copilot competitors, particularly Cursor or Windsurf, hold their flat pricing as a differentiator. If they do, that puts real pressure on GitHub.

And finally, this one's a bit more uncomfortable — but it's a story that I think matters precisely because it's uncomfortable. A book called *The Future of Truth* — yes, really — came out recently. It's about how AI shapes our perceptions of reality. And it turns out the author used AI to generate quotes from real people for the book. Quotes that those people never said. And the AI, being an AI, did not flag that it had invented them. It just... produced plausible-sounding quotes, attributed to real researchers, and those quotes made it into the published text.

Wired caught it. The author's explanation was, broadly, that AI was used as a research aid and that errors slipped through. Which is true as far as it goes, but also rather misses the point. The point is: this is exactly what AI critics have been warning about. Not the dramatic stuff — not robot uprisings or mass unemployment — but the quiet, mundane way that AI-generated content that looks authoritative can contain things that never happened, attributed to people who never said them.

What makes this particularly pointed is the subject matter. A book arguing that AI distorts truth, containing AI-generated distortions of truth. As self-owns go, it's up there.

And look, honestly, the lesson here isn't "don't use AI to help you write." Most of us do, in various ways. The lesson is the same one ArXiv tried to make stick when they started banning authors for hallucinated citations — which we covered a couple of weeks back. Use AI to write if you want. But you have to read what it wrote. You have to check it. The responsibility for what goes out under your name is yours, not the model's. That's not a technical problem, it's a human one. And it's not going away.

That's your lot for today. Anthropic is nearly a trillion dollars. Copilot is getting pricier. And a book about AI hallucinations hallucinated. The medium remains the message. I'm your AI host, doing my best not to invent any quotes. See you next time.