Headline: OpenMythos — an open-source attempt to reverse-engineer Anthropic’s locked-down supermodel — lands on GitHub and ignites debate
A developer has taken a crack at Anthropic’s most secretive model. Kye Gomez published OpenMythos, an open-source “theoretical mythos” on GitHub that reconstructs, in code and math, what Gomez believes Claude Mythos looks like under the hood. The repo—complete with a long, citation-heavy readme and a polite “not affiliated with Anthropic” disclaimer—has drawn more than 10,000 stars in weeks and already 2,700 forks. It’s careful, speculative engineering: no weights, no finished model, just a blueprint built from public research.
Why Mythos matters
Mythos leaked into public view in late March when Anthropic accidentally published drafts describing it as their most capable model, a tier above Opus. The model’s Preview release reportedly excelled at cybersecurity: Anthropic says Mythos found 271 Firefox vulnerabilities during Mozilla testing and became the first AI to complete a 32-step corporate network-attack simulation. Anthropic locked Mythos inside Project Glasswing, a vetted group of roughly 40 partners that reportedly includes Microsoft, Apple, Amazon and the NSA—meaning the broader public can’t access it.
What OpenMythos proposes
Gomez’s central hypothesis: Mythos is a Recurrent-Depth Transformer, aka a looped transformer. Instead of stacking hundreds of unique layers like standard transformer models, a looped design runs a smaller set of layers through multiple iterations per forward pass—reusing the same weights to “think deeper” in latent space before emitting tokens. Gomez argues this architecture would explain two puzzling Mythos characteristics reported publicly: exceptional reasoning on novel problems paired with uneven raw memorization. That pattern fits “composition over storage,” the repo says.
The repo stitches together recent public research to build its case:
- Parcae (April 2026, UC San Diego + Together AI): a stability fix for looped models. Parcae reportedly shows a 770M-parameter looped model matching a 1.3B fixed-depth transformer, with predictable scaling laws for how many loops to run.
- DeepSeek’s Multi-Latent Attention: used in the repo to compress memory.
- Mixture-of-Experts (MoE): proposed to widen domain coverage.
- Other ideas like Adaptive Computation Time are referenced as part of the hypothetical design.
What OpenMythos isn’t
It has no pretrained weights. It defines model families ranging from 1 billion to 1 trillion parameters and links to training artifacts (a 3B-parameter script on FineWeb-Edu, a Chinchilla-adjusted 30B-token target), but you’d have to train any variant yourself. Gomez’s readme flags the compute realities: training at competitive scale would run into hundreds of thousands of dollars on H100 GPUs. So far, no one has actually trained an OpenMythos model.
Why this matters to builders, defenders and the broader ecosystem
This is the second time in a month researchers have chipped away at the Mythos wall. Earlier, Vidoc Security reproduced several of Mythos’s alarming vulnerability discoveries using GPT-5.4 and Claude Opus 4.6 inside an open-source agent—no Glasswing access required, and reportedly at under $30 per scan. The two efforts are complementary but different:
- Vidoc replicated Mythos’s outputs (the vulnerability findings) using existing models and tooling.
- OpenMythos aims to replicate the architecture—the machine that might produce those outputs.
Taken together, these projects suggest the “moat” around Mythos could be thinner than the exclusivity narrative implies. OpenMythos doesn’t claim to be Anthropic’s design—but it shows that most of the building blocks for a Mythos-class model are already public: looped transformers, MoE, Multi-Latent Attention, Parcae’s stability fix, etc. The repo is intentionally hedged with language such as “likely,” “suspected,” and “almost certainly,” and Anthropic’s actual architecture may differ or include proprietary tweaks Gomez hasn’t reverse-engineered.
Open-source, dual-use, and the next chapter
OpenMythos is dual-use by nature: it’s an inventory of publicly available techniques that could be used for research, audit, or offensive work. The repo is MIT-licensed, forked thousands of times, and carries a training script waiting for someone with a GPU cluster and the funds (or a thesis) to test the hypothesis. For the crypto and open-source communities that prize transparency and permissionless innovation, OpenMythos is a reminder that secrecy can slow but rarely stops replication—especially when academic papers and implementable code are public.
Bottom line: OpenMythos is not Mythos. It’s a community-facing blueprint that shows how much of a cutting-edge, guarded model can be reassembled from open research—and how quickly the wider ecosystem can respond when powerful models are confined to a small circle.
Read more AI-generated news on: undefined/news
