Yesterday the federal government forced Anthropic to take its two most powerful AI models, Fable 5 and Mythos 5, offline, only days after Fable 5 reached the public. The order cited national security and required the company to cut off every customer to comply. It appears to be the first time Washington has pulled a publicly deployed AI model off the market.
It is a significant moment, and most of the coverage is missing what it actually shows. This is less a story about a dangerous technology and more a story about hype, which is becoming the most powerful force in how AI gets built, sold, and now regulated.
The threat was small. The reaction was not.
The order traced back to a method for getting around some of Fable 5's safeguards. Anthropic reviewed it, called it minor, and noted that the same capability already exists in other widely available models. The economics sharpen the point. Fable 5 is one of the most expensive AI models on the market, around fifty dollars per million tokens of output, and on real-world work it is not dramatically better than far cheaper options. Anyone trying to cause real harm with it would spend enormous sums, paid to an American company, for marginal results, long before getting anything useful out of it.
To be clear, AI does carry real risks, and government has a legitimate role in addressing them. The problem here is not oversight. It is oversight that moves on a headline before the evidence is in.
The new models are barely better than the old ones
Step back from this one model and a clearer pattern shows up. Every few weeks a lab announces the most powerful model ever built, and each release gets covered as a revolution. The improvements are real, but they are incremental. The newest flagship usually sits only a few points ahead of the one before it, and ahead of its closest competitor, on the very aggregate scores the labs choose to promote. On an actual task, the difference is often hard to notice. These systems are genuinely strong at a narrow set of hard problems, and strongest of all at writing code, but they are not the leap the marketing describes. On some measures that matter, such as how often a model simply makes things up, the most expensive flagship is not even the leader.
There is a reason the hype outruns the substance. An enormous amount of money has been poured into AI, and far too much is now riding on it to let the story cool off. The benchmark wins that fill the headlines come from expensive, carefully chosen tests, and if you spend enough chasing a particular result, you will get it. None of that tells a business much about what the technology will actually do for it.
Smaller, focused models do most of the real work
For most businesses, the practical takeaway is freeing. The most powerful model is almost never the right one. Frontier models are costly to run, they need a layer of other software built around them before they are useful at all, and they are trained on a broad sweep of the internet rather than on the realities of your business. Meanwhile, a great deal of genuinely useful work, such as summarizing documents, drafting and cleaning up writing, sorting records, and automating routine tasks, is handled just as well by smaller and far cheaper models built for a focused job.
Many of those smaller models are open source and can run on hardware a business already owns, which keeps sensitive data in the building rather than sending it to an outside vendor. Matched to the right task, a small, inexpensive model is not a compromise. It is usually the better choice, and the cheaper one.
Hype-driven policy has a track record
This is also not the first time policy has chased a headline. Years of restricting advanced chips and AI exports to China did not slow China down. They pushed China to build its own chips, to develop models that now rival the best American ones at a fraction of the cost, and to release many of them open source. The restrictions did not contain the technology. They made the competition faster, cheaper, and more independent. Reasonable people can disagree about export policy, but the pattern is hard to ignore. Rules built on the loudest story of the moment often produce the opposite of what they intend.
How we work
Cutting through all of this is the work we do. We help businesses see past the marketing to adopt AI that is practical, affordable, and matched to the actual problem in front of them, and we provide the technology leadership to run it. We also track the policy landscape that increasingly shapes it, because we work in that world directly. Because we run our own businesses, the judgment comes from inside the arena rather than from a consultant handing over a deck.
If you are weighing what AI actually means for your business, or for the rules forming around it, we are glad to talk.