The lesson from Washington’s intervention against Anthropic’s Fable 5 and Mythos 5 is not that governments are powerless over AI. They are not. A state can order a company to switch off a model. It can gate access. It can ration release to approved organisations. It can turn a commercial launch into a political permissioning process overnight.
But that is not the same as containing the capability.
That distinction matters because AI capability no longer lives only inside one model, one company, or one release. It is increasingly a moving frontier produced by falling compute costs, model-learning curves, open-weight diffusion, and rival systems constantly catching up with whatever looked unique a few months earlier.
When the US government forced Anthropic to disable Fable 5 and Mythos 5, the objective was clear: keep a dangerous vulnerability-finding capability out of hostile hands. The problem was also clear. Anthropic itself indicated that the capability in question was already obtainable from other models, including OpenAI’s GPT-5.5. The order removed a product. It did not remove the underlying ability.
That is the mechanism the intervention exposed.
AI containment through model removal runs into two forces moving in the opposite direction.
The first is falling cost. The price of delivering a fixed level of machine intelligence has collapsed. A level of model quality that cost roughly US$20 per million tokens in late 2022 cost roughly seven cents two years later. That does not mean every frontier capability can be replicated instantly or cheaply. Training, chips, data centres, power, and talent still matter. But the direction of travel is unmistakable: the cost of delivering useful machine intelligence keeps falling, and each decline lowers the barrier for competitors to reproduce more of what the frontier once made scarce.
The second force is catch-up. Frontier AI is not a single static asset. It is a moving pack. Closed models lead, open models narrow the gap, foreign models improve, in-house systems absorb specific capabilities, and fine-tuned variants spread into specialised uses. The thing that looks like a unique model capability at launch is often a broadly reproducible capability months later, sometimes sooner. The frontier moves, but the field behind it moves too.
Also Read: How AI and blockchain could make commerce decisions more accountable
Those two mechanisms together make government model bans structurally unstable.
If a government removes a model while the cost of replicating its capability is falling, the government has not raised the barrier. It has mostly raised the incentive.
It has told the market three things at once.
- First, this capability is valuable.
- Second, demand for it will remain unmet.
- Third, any supplier that can provide it outside the government’s reach now has a stronger reason to do so.
That is why the hydra metaphor fits. Cut off one model, and the intervention rewards everyone working on alternatives: open-weight developers, foreign labs, enterprise in-house teams, sovereign AI programs, and rival model companies. The ban does not erase the capability. It advertises the capability.
This is the part of AI policy that conventional debates often miss. The question is usually framed as whether a model is dangerous enough to restrict. That is a legitimate question. Some capabilities may be dangerous. Some access controls may be justified. But the harder question is what the restriction does to the market around the model.
In ordinary regulation, restricting access to a dangerous product can reduce availability. In AI, the intervention can do the opposite if the underlying capability is reproducible and the economics of reproduction are improving.
A hosted frontier model is easy for a government to reach. It sits inside a company. It is accessed through accounts, APIs, contracts, billing systems, and cloud infrastructure. That makes it controllable. It also makes it fragile. A buyer building a critical workflow around that model now has to price in political shutoff risk.
For a casual consumer, that may be an inconvenience. For an enterprise, it is different. If an AI model is embedded in software development, cybersecurity, customer operations, research workflows, or product features, a forced shutdown is not a policy event. It is an outage. A supplier that can be switched off by government order becomes a continuity risk.
That risk changes demand.
Companies will not stop wanting the capability. They will look for versions of it that cannot be removed so easily. That means more interest in open-weight models. Once an open-weight model is released, it cannot be recalled in the same way a hosted model can. It means more interest in in-house models and captive systems, where the capability is consumed internally rather than sold as third-party access. It means more interest in foreign and sovereign supply, especially from governments and companies that do not want critical AI capability dependent on another country’s permission.
Also Read: Indonesia’s AI hiring gap is real, just not 28×
Each restriction therefore sorts demand toward the forms of supply the next restriction is least able to reach.
That is the ratchet. A model can be pulled. A capability, once reproduced across more developers, more jurisdictions, more open systems, and more internal deployments, cannot easily be un-reproduced. Each intervention leaves the field more distributed than before.
This does not mean all AI controls are futile. The strongest objection is real: the bottleneck may not be the model layer. It may be the infrastructure beneath it.
Cheap tokens still require expensive chips, data centres, power, cooling, networking, and semiconductor supply chains. Those constraints are more concentrated than model access. Advanced chips and the tools to make them are physical, scarce, capital-intensive, and easier for states to govern. If the true bottleneck is compute, not model release, then governments may still be able to throttle replication by controlling chips, fabrication tools, cloud access, and power infrastructure.
That is the serious limit to the hydra argument. Model-layer bans may multiply rivals, but chip-layer controls can still slow how fast those rivals grow heads.
Even then, the policy implication changes. The effective control point is not the already-released model. It is the underlying supply chain. Pulling a commercial model after launch is the most visible form of control, but it may also be the least durable. It signals value, disrupts trusted suppliers, and pushes demand toward less controllable alternatives. Controlling compute is harder, more expensive, and geopolitically messy, but it at least targets the layer where scarcity still exists.
This is why the June interventions matter beyond Anthropic or OpenAI.
They show the early shape of a new regime. Frontier AI may no longer be treated as an ordinary commercial product. It may become a permissioned capability, released through government-reviewed access lists, rationed by user category, nationality, sector, or political approval. That may sound safer. In the short run, perhaps it is. But in the long run, a permissioned frontier creates its own counter-pressure.
Also Read: The barrier to AI adoption was never budget, it was knowing where to start
Every company that depends on AI will ask whether its supplier can be turned off.
Every foreign government will ask whether its national security systems can depend on another country’s approval.
Every developer building open alternatives will see stronger demand.
Every rival lab will see proof that the banned capability was important enough to frighten Washington.
That is the paradox. The more dramatically a government signals that a capability is too important to release, the more strongly it tells the world what to rebuild.
The state can pull a model. It can gate a launch. It can force a company to choose between compliance and continuity. But it cannot repeal the economics underneath the technology. Compute-delivered intelligence is getting cheaper. Rival models are catching up faster. Capabilities are moving from single products into distributed ecosystems.
Containment may still work where the bottleneck is physical: chips, power, data centres, semiconductor tools. It is much weaker where the bottleneck is a model that can be matched, fine-tuned, copied, approximated, or rebuilt.
That is the lesson from Fable and Mythos. Washington tried to remove a capability by removing access to a model. Instead, it gave the market a map: this capability matters, demand exists, and whoever can supply it beyond the reach of the next order will be rewarded.
That is how a ban becomes a signal.
And in a technology built on falling costs and fast catch-up, a signal can create more of the thing it was meant to suppress.
—
Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.
The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.
Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.
The post Washington banned Mythos and Fable: It created a hydra appeared first on e27.