The secretive nature of cutting-edge artificial intelligence development has once again prompted a community-driven effort to demystify a powerful, proprietary model. A developer known as Kye Gomez has released "OpenMythos" on GitHub, an ambitious open-source project aiming to reverse-engineer and reconstruct the internal architecture of Anthropic’s highly capable, yet closely guarded, AI model, Claude Mythos. This initiative, which has garnered over 10,000 GitHub stars in mere weeks, offers a detailed, code-based hypothesis about Mythos’s design, complete with extensive documentation, equations, and citations, while explicitly disclaiming any affiliation with Anthropic.
The genesis of this endeavor lies in the accidental public disclosure of information regarding Mythos in late March. Anthropic, a prominent AI safety and research company, inadvertently published draft materials describing Mythos as its most advanced model to date, positioned to surpass its Opus model in capability. A subsequent iteration, Mythos Preview, reportedly demonstrated an unprecedented proficiency in cybersecurity tasks. Official statements from Anthropic at the time highlighted Mythos’s remarkable performance during security testing, including identifying 271 vulnerabilities within the Firefox browser during evaluations conducted by Mozilla. Furthermore, it achieved a significant milestone by becoming the first AI model to successfully complete a complex, 32-step corporate network attack simulation. In response to its potent capabilities, particularly in sensitive areas like cybersecurity, Anthropic confined Mythos within "Project Glasswing," a carefully vetted consortium of approximately 40 partner organizations, including major technology players like Microsoft, Apple, and Amazon, as well as governmental bodies such as the NSA. This stringent containment strategy effectively rendered Mythos inaccessible to the public, fueling speculation and curiosity about its underlying mechanisms.
OpenMythos represents a structured response to this lack of public access. Gomez’s central hypothesis posits that Mythos employs a "Recurrent-Depth Transformer," also referred to as a "looped transformer." Unlike conventional AI models that typically stack hundreds of distinct neural network layers sequentially, looped models are theorized to utilize a smaller, fixed set of layers that are repeatedly applied in a cyclical fashion during a single forward pass. This architectural approach means that the same set of learned parameters, or "weights," are processed through multiple iterations, enabling a form of "deeper thinking" within a continuous latent space before any output, or "token," is generated.
The rationale behind this architectural guess is rooted in two observed characteristics of Mythos that have puzzled observers. Firstly, Mythos has demonstrated an exceptional ability to reason through and solve novel problems that have proven intractable for other AI models. Secondly, its raw memorization capabilities are noted to be inconsistent. The looped transformer architecture, as argued by Gomez, could explain this dichotomy. The iterative processing allows for complex compositions of learned knowledge and reasoning processes, leading to superior problem-solving in unfamiliar scenarios, while potentially leading to less consistent recall of discrete pieces of information compared to models with vastly more unique layers dedicated solely to storage. This emphasis on "composition over storage" is presented as a potential fingerprint of the looped design.
To bolster this hypothesis, OpenMythos draws upon recent advancements in AI research. The project references "Parcae," a paper published in April 2026 by researchers from the University of California San Diego and Together AI. Parcae reportedly addresses a long-standing instability issue inherent in looped model architectures. The paper details how a 770-million-parameter Parcae model can achieve comparable performance to a 1.3-billion-parameter fixed-depth transformer, while also elucidating predictable scaling laws that govern the optimal number of loops to employ. Additionally, OpenMythos incorporates elements from DeepSeek’s "Multi-Latent Attention" mechanism, designed to compress memory more efficiently, and a "Mixture-of-Experts" (MoE) setup, which allows the model to dynamically select and utilize specialized sub-networks for different tasks, thereby enhancing its breadth of knowledge and performance across diverse domains.
Despite its detailed architectural proposal, OpenMythos is fundamentally a theoretical framework. The repository does not contain the actual trained "weights" of the AI model, meaning it provides the blueprint and techniques but lacks the executed intelligence. The code defines hypothetical model variants ranging from 1 billion to 1 trillion parameters. However, to bring these theoretical constructs to life, an individual or organization would need to undertake the significant task of training these models from scratch. The project’s readme file points to a 3-billion-parameter training script intended for use with the FineWeb-Edu dataset, targeting approximately 30 billion tokens, adjusted according to Chinchilla scaling laws. Such a training endeavor would necessitate substantial computational resources, estimated to cost hundreds of thousands of dollars when utilizing high-performance hardware like NVIDIA H100 GPUs. As of now, no entity has publicly reported undertaking this massive training undertaking.
The significance of OpenMythos extends beyond its theoretical nature, particularly when considered in conjunction with other recent developments. This release marks the second instance within a short period where the perceived exclusivity surrounding Mythos’s capabilities has been challenged. Earlier in the month, Vidoc Security published a study demonstrating the replication of several of Mythos’s alarming vulnerability findings. Vidoc Security achieved this by employing publicly available models, specifically GPT-5.4 and Claude Opus 4.6, integrated within an open-source agent framework. Crucially, this replication was accomplished without any access to Anthropic’s proprietary Glasswing environment and at a cost of under $30 per scan. While Vidoc’s work focused on reproducing Mythos’s outputs – the discovered vulnerabilities – OpenMythos aims to replicate its architecture – the underlying machine that generates those outputs. The combined message from these two initiatives is potent: the "moat" protecting Mythos’s advanced capabilities may not be as insurmountable as initially suggested.
While both OpenMythos and the Vidoc replication address the mystique around Mythos, they do so through distinct methodologies. Vidoc’s research offers evidence that the results Mythos achieves, particularly in identifying security flaws, can be replicated using existing, accessible AI models. This suggests that the critical insights Mythos provides are not entirely unique to its proprietary architecture. In contrast, OpenMythos delves deeper, attempting to reconstruct the engine itself. If successful, it implies that the technological advancements that enable Mythos-class performance could, in theory, be independently developed and deployed by others.
Anthropic has not publicly commented on Kye Gomez’s architectural hypotheses, and it is highly probable that the company does not publicly share the precise design details of Mythos. Indeed, Gomez himself acknowledges the speculative nature of his work. The readme file for OpenMythos is deliberately framed with cautious language, employing terms such as "likely," "suspected," and "almost certainly." This serves as a crucial disclaimer, underscoring that the looped transformer architecture is a plausible guess, but not a definitive revelation. The actual Mythos model might not be a looped transformer at all, or it could incorporate nuanced design elements that have eluded reverse-engineering efforts thus far.
Despite these caveats, OpenMythos serves as a compelling demonstration of the maturity and accessibility of the underlying AI research. The project highlights that many of the foundational components necessary to build a model of Mythos’s caliber are already part of the public domain. Concepts such as looped transformers, Mixture-of-Experts, Multi-Latent Attention, Adaptive Computation Time, and stability fixes for looped models are not proprietary secrets. Instead, OpenMythos functions as an inventory of publicly available knowledge, demonstrating how these disparate pieces can be assembled into a coherent architectural proposal for an advanced AI system.
The OpenMythos repository is released under the MIT license, a permissive open-source license that encourages widespread adoption and modification. The project has already seen significant community engagement, evidenced by its 2,700 forks, indicating a strong interest in its potential. The presence of the training script, albeit requiring substantial resources, signifies a tangible starting point for any researchers or organizations with the necessary computational power and a research objective to pursue the development of their own Mythos-class models. This initiative underscores a broader trend in the AI landscape: the increasing democratization of advanced AI development, driven by open-source collaboration and the rapid dissemination of research findings, even as leading companies strive to maintain proprietary advantages in their most potent creations. The ongoing efforts to dissect and replicate "Mythos" not only illuminate the technical underpinnings of advanced AI but also raise important questions about the future of AI safety, access, and competition.
