Well, Fable is back. Nerfed, obviously. Let me walk you through the timeline of this whole hype circus around Mythos and then Fable, because we have seen this movie before, more than once, and every single time we fall for it. Every time the hype, every time the nonsense.
We have seen this before: GPT-2
This pattern is not new. Think back to GPT-2. At some point OpenAI came out and said: “We built GPT-2, but we can’t release it, because it could be abused for spam and it’s a huge risk. Oh no, we’re all going to die.”
A month later they released it, quietly. People started using it, and it was a model that was more or less fine syntactically, but semantically it was all over the place. They were not usable chatbots yet.
With Mythos the same thing happened, only much longer, much more intense, much more interesting, and much more tied to a theme that fascinates me: surveillance.
The Mythos saga
So Anthropic comes out and says: “We’ve built the most powerful model in history. It’s called Mythos, but we can’t release it because it’s too dangerous.” And yet they tell everyone, and they publish a 200-page report where they explain that they had a psychiatrist analyze Mythos, and then Mythos escaped the sandbox and emailed a researcher while the researcher was calmly sitting in the park eating a sandwich.
So: the most important model in history, the most powerful, but too dangerous. Here’s the detailed report on how it works. But we won’t give it to you.
Then they announced Project Glasswing. With Project Glasswing, they said, they wanted to give early access to this super-powerful, super-dangerous model only to a restricted circle of humanity: the part of humanity they can trust, the part that truly cares about the well-being of the planet and its people, the part that keeps us safe, the part that represents the citizenry.
In practice, they gave early access to Mythos to all their friends and their biggest B2B clients. They threw in the Linux Foundation, just to signal that there’s something non-profit in the mix, so its presence could “balance out” the likes of JPMorgan Chase in the room. Because the worst had to be counterbalanced somehow.
The researchers push back
Meanwhile, the few researchers who actually had this otherwise-unusable Fable in their hands started saying: sure, it may be more powerful, but if you take something like Opus 4.7 and run it long enough - leaning on test-time compute, making it produce more tokens - it gets to the same vulnerabilities anyway. It’s not that only Mythos can do it and the other models can’t. But by then it’s too late for the truth to matter.
And in my own small corner, this bald guy here has said it more than once, and I’ll repeat it: this is nonsense. This is nonsense.
Guardrails and the shifting excuse
After a while they come back with Fable, which is basically a version of Mythos with guardrails. These guardrails are meant to stop the model from being used for biological harm and for cybersecurity - for hacking.
Notice how the excuse keeps rotating. First it was spam. Now it’s the biological problem, the security problem. Then at some point they’ll add something else. Always the same nonsense.
Enter the government
Then the American government steps in and says: “It’s too dangerous, we agree with you, so everybody freeze, block everything, because only American citizens can use this.”
And what does citizenship and nationality have to do with any of this? It matters because, a few days after the model got blocked, Anthropic started asking for ID documents. And people around the world started talking about the nationalization of these models.
This slots neatly into a path that’s already underway: Digital ID, Central Bank Digital Currency, the steady tightening of web services. It all falls beautifully into the same big trend we’ve been watching, and it can only keep going that way.
Nerfing models is not new either
Let me add a few words about the nerfing of these models. This is not new either. This morning I saw statistics on the latest Fable release, and it’s genuinely worse than the Fable they published just a few weeks ago - and nobody really knows how any of it relates back to Mythos, which is, fittingly, a myth.
Making models dumber has happened again and again. I remember it well, and we noticed it in our own small Cheshire Cat community. We were using GPT-3 - I had access to the GPT-3 API even before ChatGPT came out. Back then, in the Cheshire Cat, we did the tool parsing by hand. There were no tool-calling APIs, so it was all done inside the Cat with a prompt that parsed the output. An insane setup, if you think about it, and one I’m still quite proud of.
Then, right after ChatGPT blew up, both ChatGPT and GPT-3 suddenly - within a few days - could no longer fire the tools in our tests. Things they had been doing just days earlier, in particular using tools, stopped working.
I complained about it. In the community there was a guy from Microsoft, and he publicly challenged me: “You can’t talk like that. You have no proof. These are serious companies.” Microsoft at the time was investing enormously in OpenAI and in blasting AI in every direction - they’d opened their own APIs, they had exclusivity on OpenAI’s models. He told me I had no right to say the models had been nerfed.
A few weeks later, a paper came out from people who had tracked these models’ performance over time behind the American APIs, and they showed that both ChatGPT and GPT-3 had indeed suddenly become less capable on benchmarks.
The key: it’s the compute
Why does this happen? And I apologize to my long-time followers for repeating myself so much, but the essential lens for reading all of this is compute - energy and compute.
Running this stuff takes a lot of machines, it costs a lot, and the low prices they’ve kept so far - subscriptions, even letting hundreds of millions of people use ChatGPT and friends for free - is all money on credit. It costs far more than what they charge, and they don’t have enough compute to serve everyone who wants it. So they’re squeezed on both ends at once: using fewer and fewer resources while charging more and more for the resources they do use.
So what do they do? They publish the super-model at maximum quality, let people try it, and then they put the R&D teams to work building models that are distilled or quantized or otherwise made more efficient - roughly the same performance, but cheaper to run.
Why I care
And why does all this matter to me? Because in the coming period I’ve given myself a new mission: to pick which models to actually build on. I’m looking at all of them - Fable, Mythos, and the Chinese models too - and I have to choose.
Same story, over and over. How many more times do we have to watch this? Probably another fifty. We keep lining up like sheep. But keep your eyes open: the pattern is always the same, and knowing it is what lets you see through the next round of hype.