AI companies also cry
Drama, betrayal, and billions of dollars: Anthropic, OpenAI, the US government, and the Chinese models in the telenovela you can’t afford to miss.
A couple of weeks ago, Anthropic released its most powerful public AI model, Claude Fable, and three days later they had to pull it from the market at the request of the US government over “national security” concerns.
Yes, this model was only available for three days.
All of this inspired me to tell you this “telenovela”, starring Sam Altman, Dario Amodei, a few Chinese AI companies, Donald J. Trump, and Germán Martínez (as himself) 😅.
Let’s dive in!
Note: as telenovela-flavored as this post is, this is an important topic. It’s about the fact that we don’t just depend on AI companies, we’re also at the mercy of governments to use these models. And that’s concerning.
An AI too dangerous to release
This April, Anthropic announced they had a model so powerful they were a little scared to just release it. They said Mythos (that’s the model’s name) was extremely good at hacking software and could find a gazillion security holes in all kinds of programs, so they preferred to give it first to a few companies so they could “patch” their systems before releasing it to the public.
Here’s where the timing of the announcement looks a little suspicious, because the people at Anthropic were preparing to go public on the stock market (an IPO, as they call it). And as we all know, a stock’s value doesn’t depend so much on what a company does, but on what the market believes that company can do in the future. So announcing that you have the most advanced model in the whole wide world (and that it’s too dangerous to give us) couldn’t have come at a better time.
So, to keep the hype going, Anthropic announced Fable 5, a “friendlier” version of Mythos. Deep down it has the same capabilities, but if it received an instruction that seemed dangerous, it automatically passed the question to a less powerful model, like Opus 4.8.
Everyone went crazy and we all ran to try it (including me). I started with a quick test of conversation and code generation... and, to be honest, my first impression was “meh!”. So I thought I’d come back a couple of days later to run some heavier tests. The problem was that after three days, the US government “kindly asked” them to pull it from the market over national security concerns... CHAAAAAAN!
But this story is just getting started...
What happened here?
What happened was that only two days after its launch, someone managed to get around Fable 5’s security measures and convinced it to answer the questions it wasn’t supposed to. So the whole “it’s the same as Mythos, but we have it under control” thing wasn’t as true as they thought.
Note: you remember that this is how 89.7% of apocalyptic science fiction movies start, right?
So, when the US government found out the model was no longer under control, they demanded that Anthropic shut it down. To be more precise, they didn’t ask them to shut it down, but to make sure no foreigner could use it. And since they had no way of knowing the nationality of each of their users, they ended up shutting down everything.
And here’s where the gossip begins! Several months ago, Anthropic had a fight with the Pentagon because they didn’t allow Claude to be used for autonomous weapons applications or mass surveillance of (US) citizens. The government’s response was to declare them a “national security risk” and cancel a few hundred million dollars in contracts they had with them.
Rumor has it that this move with Fable could have been a kind of retaliation for their refusal to give them Claude without restrictions. Although these are just rumors.
But then, a few weeks later, the same thing happened to OpenAI, which launched ChatGPT-5.6 with limits (also at the request of the US government), and only for a group of carefully selected companies.
It’s a strange pattern that two of the biggest AI companies in the world, both about to go public on the stock market, say almost at the same time that they have the “most powerful model in the world”, so powerful that their government wants it to be used only by their fellow countrymen. Something doesn’t add up here...
Was it reeeeally that dangerous?
I don’t know, but just a few days ago Anthropic announced that Fable 5, that model so destructive it had to be shut down across the entire planet, was available again. ¯\_(ツ)_/¯
The big security solution? They added an extra filter to block the trick that had been used to hack it. And with just that, it stopped being a threat to the human race (again, this is how Terminator started)...
And today, July 9 (as I was writing this post), OpenAI announced that its model was also going public, after the US government ran some tests and approved it.
In other words, the two models that were “too dangerous to release” will both be on sale within a couple of weeks for anyone to use.
Let’s look at the full pattern for a second (this is pure speculation, but here it is):
AI company wants to go public
It announces it has a super advanced model that is extremely dangerous.
The US government steps in and bans us from using it (worldwide drama)
A few weeks later the model comes back (with all the hype)
The stock value at the IPO goes through the roof.
And now that we can finally try the new models....
I finally got to try Fable 5
My verdict? It’s a very good model, but it burns tokens like you wouldn’t believe.
These past few weeks I’ve been using Claude CoWork and Code a lot (posts about that are coming), so I decided to try Fable 5 with those tools. The result, in my opinion, was a bit better than what Opus 4.8 would have given me, but I spent all the tokens I had available for 5 hours in 25 minutes, and that’s with the Max plan, which gives me 5 times the limits of the Pro plan.
Let me say it one more time:
Fable 5 spent my entire Claude Max session’s tokens in 25 minutes.
And since I’m not a millionaire (and Substack won’t let me turn on paid subscriptions), I went back to Opus 4.8.
Oh, and by the way, starting July 13 this model will no longer be included with the Pro or Max subscriptions, and anyone who wants to use it will have to buy additional credits on top of their subscription.
Meanwhile, in China
While we were all watching this telenovela of American models, on the other side of the world, Chinese AI labs like DeepSeek, Alibaba (with its model Qwen) and Zhipu (with GLM) were releasing very powerful, open models.
What does “open” mean? That you can download them and run them on your own hardware. No fear of someone cutting off your access. Oh, and on top of that, if you don’t want to install them and decide to use them through the provider, they’re much cheaper than the OpenAI and Anthropic models.
It’s true that they’re not as good as ChatGPT 5.6 or Fable 5 yet, but they’re very, very, very close; so for most use cases you can imagine, they’re probably worth at least considering.
Another detail in this telenovela is that the US bans the export of Nvidia’s most powerful chips (Nvidia being the leading company in AI processors) to China. But even so, through pure ingenuity, the Chinese models have managed to stay almost at the same level.
Imagine where they’d be if they had access to those processors!
American companies accuse the Chinese ones
As if there wasn’t enough conflict in this novela, Anthropic and OpenAI are saying the Chinese models are a kind of “copy” of their models. Actually, they don’t say copy: they gave it a fancy name and call it distillation.
Distillation is using one model to train another. In simple terms, it’s asking the most advanced models hundreds of thousands of questions, writing down how they respond, and using that data to train your model. Which means that to train the Chinese models, these companies created thousands of accounts, paid to use Claude or ChatGPT, and used that data to catch up with them (well done, China).
Irony of ironies: Anthropic, the same company complaining that its model was used for this, reached a 1.5 billion dollar settlement in September 2025 for having trained its models on copyrighted material (ooops).
We are all part of this telenovela
Jokes aside, what should worry us here is that, even if we pay for them, we have no guaranteed access to these tools. We’re at the mercy of whoever is in government and their whims.
Imagine that one of these days you find out you can’t use an AI because you’re not a citizen of this or that country. That’s the real problem.
Now imagine that in one country (or a couple of countries), their students, their governments, their companies, their intelligence services, their banks, etc. have access to the most advanced AI models while the rest of the world doesn’t.
That’s a picture of the world I don’t like, but we can’t go around pretending it’s not a possibility.
And like every telenovela, this one has about 9,546 more episodes to go, so we’ll have to stay tuned.
All the best from Lima - Perú, see you in the next one.
G