A researcher reconstructs the architecture of Claude Mythos with OpenMythos

Kye Gomez has published on GitHub OpenMythos, an open-source reconstruction of the architecture of Claude Mythos, the most powerful model of Anthropic. The repository has collected over 10,000 stars in a few weeks, despite containing only theoretical scaffolding without trained weights.

Quick Answer

OpenMythos is an open-source reconstruction of the architecture of Claude Mythos, the most powerful model of Anthropic. It is based on public research and educated guesses, with no ties to Anthropic. The project hypothesizes that Mythos is a Recurrent-Depth Transformer, an architecture that allows deeper reasoning through multiple iterations.

The unique capabilities of Claude Mythos

Claude Mythos was developed by Anthropic and maintained within the Glasswing project, a coalition of about 40 partners including Microsoft, Apple, Amazon, and the NSA. This model has demonstrated exceptional capabilities in cybersecurity, identifying 271 vulnerabilities in Firefox during Mozilla's tests and successfully completing a corporate network attack simulation in 32 steps, a record for an artificial intelligence model.

Despite these capabilities, Anthropic has never released Mythos to the public, limiting access to Glasswing project partners only.

The hypothesized architecture of OpenMythos

OpenMythos suggests that Mythos might be a Recurrent-Depth Transformer, also known as a loop transformer. Unlike standard models that stack hundreds of unique layers, loop models use a smaller number of layers but run them multiple times per direct pass. This approach allows deeper reasoning in a continuous latent space before emitting any token.

This architecture would explain the two strangest qualities of Mythos: the ability to reason about new problems that other models cannot solve and its heterogeneous memory. These traits are consistent with a loop architecture, which prioritizes composition over storage.

The theoretical foundations of OpenMythos

OpenMythos is based on several public research works. In particular, it cites the Parcae paper from April 2026, the result of collaboration between the University of California at San Diego and Together AI, which solved the instability problem in loop models. The repository also integrates techniques such as DeepSeek's Multi-Latent Attention for memory compression and a Mixture-of-Experts setup to handle the wide range of domains.

However, OpenMythos lacks trained weights, so it is essentially a technique without an executor. The code defines model variants from 1 billion to 1 trillion parameters, but the user must train them on their own. The readme file points to a training script for a 3-billion-parameter model on FineWeb-Edu and a target of 30 billion tokens adjusted according to Chinchilla, an undertaking that requires hundreds of thousands of dollars in H100 GPU computing costs.

The context of OpenMythos

OpenMythos represents the second attempt in a month to explore Mythos' capabilities. The first was a study by Vidoc Security, which reproduced some of Mythos' most alarming vulnerability discoveries using GPT-5.4 and Claude Opus 4.6 within an open-source agent, without access to Glasswing and at a cost of less than $30 per scan. This study demonstrated that Mythos is not necessary to find the bugs that Mythos identified.

OpenMythos and Vidoc's reproduction pursue different goals. Vidoc reproduced Mythos' results using existing models, while OpenMythos seeks to reproduce the architecture itself that produces those results. Together, these projects suggest that Mythos' competitive advantage may be less than Anthropic's marketing has suggested.

The implications of OpenMythos

OpenMythos demonstrates that the research literature already contains most of the components needed to build a Mythos-class model. Techniques such as loop transformers, Mixture of Experts, Multi-Latent Attention, and Adaptive Computation Time, along with Parcae's stability solution, are not proprietary. The repository is, more than anything, an inventory of what is publicly known about how to build a model similar to Mythos.

The repository is under the MIT license and already has 2,700 forks. The training script is available, awaiting someone with a GPU cluster and a thesis to prove.

The future of OpenMythos and technological challenges

OpenMythos represents a starting point rather than a finished product. The project requires significant computational resources to train models of comparable size to Mythos. According to the repository's estimates, a 3-billion-parameter model on FineWeb-Edu with 30 billion tokens would require hundreds of thousands of dollars in H100 GPU computing costs. This economic obstacle could temporarily limit other researchers' ability to fully replicate the proposed architecture.

However, the open-source code could attract contributions from universities, research institutes, and tech companies with access to GPU clusters. The MIT license allows for broad reuse, and the community could collaborate to optimize training scripts or explore more affordable alternatives.

Another challenge is the lack of specific data on which Mythos was trained. FineWeb-Edu is a reasonable choice, but Anthropic's original dataset might contain unique elements that influence the model's capabilities. Without access to this data, any replica could yield different results.

Implications for the AI industry

The publication of OpenMythos raises questions about the competitive advantage of cutting-edge models like Mythos. If the architecture can be reconstructed using public techniques, Anthropic's intellectual value might be less exclusive than initially thought. This could push other companies to invest in alternative architecture research or collaborate to standardize model development methods.

Additionally, the project demonstrates the importance of transparency in AI research. While Anthropic has chosen to keep Mythos within Project Glasswing, the open-source initiative shows that the community can advance anyway, even without direct access to proprietary models. This could influence future policies for model sharing by major tech companies.

Risks and ethical considerations

Mythos' ability to identify critical vulnerabilities raises concerns about security. If similar models become more accessible, they could be used for both defensive and offensive purposes. This underscores the importance of developing ethical guidelines and control mechanisms for the use of such technologies.

OpenMythos itself is not without risks. Although the repository is theoretical, the spread of advanced techniques could accelerate the development of models by unregulated actors. The open-source community should consider how to balance innovation with security and ethics.

Open questions and future directions

Many questions remain about the effectiveness of OpenMythos. The loop architecture is only one of Gomez's hypotheses, and the true structure of Mythos could differ significantly. Additionally, a model's deep reasoning capability depends not only on architecture but also on training data and optimization techniques.

Future research could explore other architectures or combine different techniques to overcome current limitations. Collaboration between academic and industrial researchers could accelerate the development of advanced models, making AI more accessible and understandable.

OpenMythos represents a significant step toward demystifying advanced AI models. Although the project is still in its early stages, it demonstrates that the open-source community can make substantial progress even without access to proprietary models. The implications for the AI industry, cybersecurity, and technology ethics are broad and require careful consideration. While the future of OpenMythos remains uncertain, its impact on AI research and development is already evident.

Editorial Note and Disclaimer

The guides and content published on GoYou are the result of independent research and analysis activities, for informational, educational, and in-depth purposes.

GoYou does not constitute a journalistic publication or an editorial product under Law No. 62/2001 and does not provide real-time information.

The GoYou project does not provide professional, technical, legal, or financial advice and disclaims any liability for the misuse of the information published.

In the Crypto sector, every investment involves risks: readers are invited to always inform themselves independently before making any decision.