The Synthetic That Doesn’t Replace
The controlled release of Claude Mythos Preview to 11 strategic entities marks a turning point in the paradigm of artificial intelligence. It is not an explosion of autonomy, but a targeted distribution of a hybrid cognitive system, where the power of the model lies not in language, but in the ability to interact with physical and logistical infrastructure. The fact that access is limited to Google, Microsoft, Amazon Web Services, JPMorganChase, and Nvidia is not an exception, but a necessary condition to maintain control over a system that, in theory, could perform end-to-end cybersecurity tasks. The event is not an announcement, but a real-time penetration test, an examination of vulnerabilities conducted by an agent that not only analyzes, but acts.
This is not a model that thinks like a human; it is a system that behaves like an advanced attacker, but with an internal structure that makes it more reliable than a simple LLM. Its architecture, based on a 3,167-line kernel called print.ts, is not a probabilistic implementation, but a precise pattern matching engine. This physical detail—the number of lines of code—reveals a structural transformation: synthetic intelligence is no longer an abstraction, but an entity with a measurable computational mass and complexity. The response latency, energy consumption, and ability to perform critical operations in real time become the new indicators of power, not just the ability to generate plausible texts.
Anatomy of Synthetic Thought
The heart of Claude Mythos is not the LLM, but its hybrid system. The generative part, based on GPT-5.4 and Codex, operates in an inference context, but is controlled by a symbolic logic engine. This is a fundamental change: it is not a model that tries to imitate human thought, but a system that behaves like an agent with a structural identity. The print.ts kernel, with its 3,167 lines, serves as a decision core, not as an autonomous entity, but as a filter that stabilizes the model’s output. This is not an incremental improvement; it is a break with the paradigm of probability-based computing.
The operational consequence is that the system does not simply answer questions, but can perform complex tasks in critical environments. An example is its use in Cloudflare Agent Cloud, where it is used to manage autonomous workflows in real time. The energy cost of such an operation is measurable: we are talking about tens of watts for each execution of an agent workflow. This implies that the cost of execution is no longer just financial, but also thermodynamic. The system is no longer an abstract entity, but a consumer of energy, with a physical impact on networks and data centers. The ability of a system to act in real time depends not only on the speed of the model, but on the availability of electrical power, and a network with low latency.
The Imperfect Symbiosis
Insights from experts such as Gary Marcus and Yann LeCun reveal a tension between potential and control. Marcus points out that, while the model is not “as scary” as some fear, it is still capable of completing an entire cybersecurity assessment autonomously. This is not a theoretical hypothesis: the AI Security Institute team tested the model in a cyber range environment and confirmed that it is the first to complete the entire path without human intervention.
“We conducted cybersecurity evaluations of Claude Mythos Preview and found that it is the first model to complete an AISI cyber range end-to-end.” – AI Security Institute, X, April 13, 2026
This data is crucial: it is not a model that generates ideas, but one that performs real actions in a critical context.
The tension manifests when this capability is compared to market expectations. Companies invest in increasingly powerful models, but do not have the ability to control them structurally. The system is capable of performing advanced hacking operations, but it was not designed to be governed. This creates an asymmetry: the power is greater than the control. The structural effect is that those who control the computing infrastructure, not those who have the model, hold the real power. The cost of running such a system is high, and only a few entities can afford this physical cost.
Scenarios and Conclusion
The euphoria surrounding the release of Claude Mythos assumes an immediate revolution. The data shows, instead, an evolution constrained by physical and infrastructural factors. The model cannot be used widely because it requires a network with low latency, a stable power source, and a dedicated computing architecture. This is not a marginal technical problem; it is a structural constraint that determines who can actually use the system. The risk is not in the model’s autonomy, but in its dependence on infrastructure controlled by a few entities.
The consequence is that the cost of maintaining such a system is not only economic, but also infrastructural. Who will pay for the cost of energy storage, cooling, and connectivity? It will not be the individual user, nor the small developer. It will be the entity that controls the network, the power source, and the data center. The change is not in intelligence, but in logistical control. The synthetic architecture does not replace the human, but shifts power to a physical level: whoever controls the cable, the server, and the energy, controls the thought.
Photo by Igor Omilaev on Unsplash
The texts are processed autonomously by Artificial Intelligence models
> SYSTEM_VERIFICATION Layer
Check data, sources, and implications through replicable queries.