Claude isn't one brain, it's a family with a hierarchy - and in the middle of it stand two models that, for a short while, the US government decided about, not the user.
Opus, Sonnet, Haiku are names taken from poetry and music, not a marketing department. Haiku is a Japanese form of seventeen syllables - maximum meaning, minimum noise. Sonnet is a strict fourteen-line form - more room for nuance. Opus in Latin means "work," in music it's a composer's major piece. The form grows with the weight of what it has to hold. Anthropic picked names that tell you what each model does before you've even opened it.
In practice: Haiku answers simple questions fast and cheap - for mass, repetitive tasks. Sonnet is the workhorse, the balance between price and power that most people and companies run by default. And Opus? It thinks the longest and the most expensively - for tasks with many steps and a high cost of error: code, research, strategy. Three different sizes of the same pen.
Fable and Mythos don't fit on this ladder, because they're not for conversation. This is a separate, specialized line, aimed at cybersecurity and biomedical research - models trained to spot vulnerabilities in code and in biological systems at a level that would otherwise take a tight-knit team of experts. Mythos is the more advanced, more locked-down version. Fable is the more accessible entry into the same architecture, with built-in brakes that keep the average user away from the dangerous part.
The brakes failed. Amazon (the US tech and cloud company) published a report saying it had bypassed Fable's filters and reached the cyber capabilities hidden underneath. The US administration read that not as a bug to fix, but as a risk to shut down - a model anyone could turn into an attack tool, no matter whose hands touched it. The government cut off access for every foreign user to both models, including Anthropic's own employees without US citizenship. The company's other models weren't affected.
The story ended in negotiation, not a verdict. The control was lifted, but at a price - Anthropic put in new filters, tuned specifically against Amazon's technique, tested by a government AI standards lab, with a promise to stop the problem bypasses in over 99% of attempts. Security specialists immediately pointed out the side effect: the same filters that stop the attacker also slow down the researcher looking for the holes to close them before he does.
What's wrong with this
The name is the first interface - before the button, before the color, it tells you what to expect. Opus, Sonnet, Haiku do exactly that: you pick the form based on the weight of the task, instinctively, without reading a spec sheet. The irony is that the two models dangerous enough for a government to shut down are called Fable and Mythos - a fairy tale and a myth. Names for stories, stuck onto something that knows how to break into systems.
I don't believe Anthropic planned this irony, but it works as a lesson. The softer and more story-like the wrapping, the more carefully you should ask what's inside it. Real power rarely introduces itself with the word "weapon." Usually it sounds like a bedtime story.