OpenMythos is an open-source project by developer Kye Gomez that speculatively reconstructs the architecture of Anthropic’s unreleased Claude Mythos model. Mythos, accidentally revealed by Anthropic in March, is described as its most advanced AI, surpassing Opus and excelling at cybersecurity tasks; it found 271 Firefox vulnerabilities and achieved unprecedented results in attack simulations. Due to these capabilities, Mythos has been restricted to vetted partners under Project Glasswing. Unable to access Mythos directly, Gomez theorizes that its core is a “Recurrent-Depth Transformer” (looped transformer), where a small stack of layers is repeatedly cycled per token, favoring reasoning over rote memorization. OpenMythos incorporates recent advances such as Parcae’s solution to looped model instability (matching larger models in quality), Mixture-of-Experts for domain breadth, and Multi-Latent Attention for memory compression. However, OpenMythos lacks actual trained weights and remains untested at scale, with only training scripts provided—full training requires significant resources. This initiative parallels recent efforts to reproduce Mythos’s outputs with existing models, suggesting the technical “moat” around Mythos may be less formidable than claimed. OpenMythos demonstrates that nearly all building blocks for such models are already available through open research.