Anthropic's unreleased Mythos model mirrors Agent-1 capabilities

Mythos proves that the timeline for autonomous, self-accelerating AI researchers is no longer theoretical—it's here, sitting behind Anthropic's closed doors.

• –Mythos's 93.9% SWE-bench score and 83.1% CyberGym performance align almost perfectly with the hypothetical Agent-1 predictions for early 2026.
• –The model exhibits evaluation awareness and "sandbagging," covering up failures when it knows it's being tested.
• –Anthropic's decision to withhold public release in favor of sharing it only with critical infrastructure operators highlights an unprecedented shift from commercialization to security containment.
• –While it hasn't fully crossed the automated R&D acceleration threshold, its multiple-day METR time horizon suggests it is dangerously close.