OpenMythos: Open-Source Implementation Reverse-Engineering Claude Mythos Released

OpenMythos is an open-source project attempting to recreate the structure of Claude Mythos, rumored to be Anthropic's next-generation model, based on public information and research ideas.

It is not an actual replication of the model, but rather closer to a hypothetical implementation that codifies "the possibility that it might work this way."

The core idea is that instead of making the model larger, it repeatedly executes the same Transformer block multiple times to progressively update internal state.

• Repeatedly calling the same block to induce deeper reasoning
• Selectively activating MoE (expert) during the repetition process
• Processing intermediate results within internal state rather than immediately outputting them as external tokens
• Applying an attention structure aimed at improving memory efficiency

This approach is closer to a form that computes sufficiently internally before outputting only the final result, rather than revealing reasoning while generating tokens like typical LLMs. In other words, it aims for a structure of "repeatedly computing internally before speaking" rather than "thinking while speaking."

The meaningful point is that the axis of performance improvement lies not in increasing the number of parameters but in the amount of repeated computation during the inference stage. It can also reduce the number of externally generated tokens, connecting it to the issue of token cost.

However, it has not been confirmed that this is identical to the actual structure of Claude Mythos, and verified large-scale experimental results or performance evidence are still lacking.