Over the weekend, the most powerful AI model ever built came back online. Not for you.
The US government notified Anthropic that Mythos 5, its strongest cybersecurity model, could be redeployed, but only to a set of US organizations that operate and defend critical infrastructure. A letter from Commerce Secretary Lutnick cleared it for roughly a hundred hand-picked institutions, Fortune 500 names and federal agencies, on a list the public cannot see. Fable 5, the version meant for everyone else, is still dark.
Sit with that split, because it is the whole game in miniature.
Two weeks earlier, on a Friday at 5:21 p.m. Eastern, the Commerce Department had ordered Anthropic to cut off foreign access to Fable 5 and Mythos 5, and the company pulled both worldwide within hours. The reason was simple: Mythos is dangerous. It can find flaws in code long believed unbreakable, which means a single prompt in the wrong hands could do enormous damage. That capability is exactly why Anthropic built it, and exactly why it was rationed to a vetted few in the first place.
But notice what the public got instead. Fable is the same underlying model as Mythos, with the dangerous capabilities boxed off by safeguards. The elite get the version that can act. You get the version that can talk. Same engine, different collar.
And what is that engine actually doing when it earns the label “most powerful”? It is not writing better sentences. Fable was built for long-running, autonomous work, executing complex coding and knowledge tasks for extended stretches with no human in the loop. It reads an entire codebase, reasons through thousands of paths, holds the whole structure in memory, and keeps going long after a person would have stopped. That is the frontier everyone is sprinting toward. It is also, by no coincidence, the single most token-hungry thing a model can do.
This is the bait.
The industry has names for it: reasoning, thinking, agents. It is sold as the largest leap since the chatbot, and in raw capability the claim holds. But watch where the cost lands. A reasoning model can require more than a hundred times the compute of a plain answer. The headline price per token hides the real cost by five to thirty times, depending on the task. And the tokens doing the damage are the ones you never see. The model mutters to itself before it speaks, and you are billed for the muttering.
That is what makes it the sweetest bait of all. The old token tricks were crude: a padded summary, a longer answer, things you could at least notice. Reasoning tokens are invisible by design. When a model spends nine thousand tokens thinking before it hands you four hundred, you cannot audit whether eight thousand of those were necessary or were simply the meter running.
Now watch the tell.
The Mythos restart landed just hours after OpenAI said it would release its new GPT-5.6 models only in stages, at the government’s request, instead of the wide launch it had planned. Sam Altman called the staggered rollout “bad news.”
Read that again. A safety brake gets applied to the most capable systems on earth, and the reflex from the top of the industry is not relief. It is disappointment. Wide release was the plan, and wide release is the business, because wide release is volume, and volume is the only thing the meter cares about. When the people building this technology are told to slow its spread, they experience it as a loss. That reaction is the entire thesis of this series, captured in two words.
There is a local epilogue worth noting, because it is not abstract for readers here. The original shutdown reportedly traced, in part, to a Korean telecommunications company with access to Mythos and suspected ties to China. The most powerful model in the world, the question of who may touch it, and a Korean node at the center of the blast radius. The trap is not a Silicon Valley curiosity. It runs through Seoul.
So the bait has evolved. It is no longer a longer answer. It is power, delivered through invisible token-burning work, rationed to a list you are not on, and strained against every brake the world tries to set. We lunge for it anyway, never asking who holds the other end of the line.
Next, Part 3: the trap’s quietest mechanism. The price per token keeps falling, and your bill keeps climbing. How both are true at once, and why the falling price is the hook, not the mercy.
