Claude Mythos: Just abliterated Opus 4.6?

The theory that Claude Mythos is just an "abliterated" version of Opus 4.6 highlights the massive "alignment tax" currently dampening frontier model capabilities. Mythos’s 100% Cybench score suggests capabilities that go beyond simple refusal removal, potentially involving autonomous zero-day discovery. Local LLM users report significantly better performance on cybersecurity tasks using abliterated Opus 4.6 fine-tunes. Anthropic’s restriction of Mythos to "Project Glasswing" partners suggests the model’s social and political risk remains the primary barrier to public release. This implies that frontier capability leaps are increasingly being achieved by removing safety bottlenecks rather than architectural breakthroughs.