METR's Claude Mythos benchmark shows AI progress on coding tasks, but analyst cautions against extrapolating to general superintelligence

METR evaluated an early version of Claude Mythos Preview in March 2026 and estimated a 50%-time-horizon of at least 16hrs (95% CI 8.5hrs to 55hrs) on its task suite, extending the upper end of what the think tank can measure without new tasks.

The 50% success threshold is arbitrary and leaves headroom if higher performance levels (80% or 95%) are demanded; the graph measures only software-development tasks, not general intelligence, and does not address the key problem of AI reliability.

Recent advances likely stem from incorporation of symbolic tools (code interpreters, verification, harnesses) rather than model scaling alone; Mythos tracks near trend on the broader ECI benchmark and does not show acceleration beyond prior models.