OpenMythos repo reconstructs Anthropic’s Claude Mythos

Kye Gomez published OpenMythos, an open-source GitHub repository that lays out a reconstruction of Anthropic’s Claude Mythos architecture. The project includes model code, training scripts and documentation, and is released under the MIT license. The repository had more than 10,000 stars and roughly 2,700 forks within weeks of its release.

The repo presents a technical outline and an extensive readme with equations and citations. It defines model variants ranging from about 1 billion to 1 trillion parameters, points to dataset scripts and provides training recipes. The repository does not include trained model weights; any training must be carried out by users with appropriate compute resources.

Central to the reconstruction is the hypothesis that Mythos uses a Recurrent-Depth Transformer, also called a looped transformer. In a looped design a smaller set of weights is reused through multiple iterations inside a single forward pass instead of stacking many distinct layers. The readme argues that repeated iterations of the same weights could produce strong multi-step reasoning while producing uneven raw memorization.

OpenMythos cites recent research papers as building blocks. The repo references Parcae, an April 2026 paper from researchers at the University of California, San Diego and Together AI, which addressed instability in looped models. The codebase also draws on Multi-Latent Attention for compressed memory and a Mixture-of-Experts arrangement to broaden domain coverage. The documentation recommends a Chinchilla-style training target of roughly 30 billion tokens for a 3 billion-parameter run and links to a FineWeb-Edu dataset script. The readme notes that training at scale would require substantial compute, potentially costing hundreds of thousands of dollars on clusters of H100 GPUs.

OpenMythos appears after internal materials about Claude Mythos entered public view in March. Those materials described Mythos as Anthropic’s most capable model and reported that a Mythos preview discovered 271 vulnerabilities during testing of the Firefox browser and completed a simulated 32-step corporate network attack. Anthropic placed Mythos inside Project Glasswing, a vetted program of about 40 partners that includes major technology companies and government entities. The company has not publicly released access to the model.

A separate security research effort replicated several of the vulnerability findings attributed to Mythos by using off-the-shelf large models and open-source agents. That work reported similar scanning results at a lower cost and without access to the confined Mythos system.

The OpenMythos repository includes repeated caveats about uncertainty. The readme uses terms such as “likely” and “suspected” to qualify design choices and notes that the real Mythos system may use different components or proprietary tweaks. The codebase functions as an engineering blueprint and a catalog of public methods; it provides a path for groups with large compute budgets to attempt full training but does not supply a trained model.

Until weights are trained and released, OpenMythos remains a speculative reconstruction built from public research and informed conjecture. The repository maps public techniques that could be combined to create a system with characteristics similar to those described in the leaked materials.