// TAG

#metr

A progress curve bending sharply upward after a point labelled AI helps, beside a report page with most lines blacked out

METR finds Anthropic's AI is now speeding up its own development, and Anthropic redacted most of the evidence

After 11 weeks inside Anthropic, METR says its models now improve about twice as fast as a year ago, mostly thanks to AI. Staff report being 12 times as productive.

BREAKING Nov 8, 2026 · 2 min

A human reviewer behind a towering pile of alert cards, a few marked real in pink and most marked false in gray

// AI vs Humans

AI auditors catch most agent cheating, but bury reviewers in false alarms

A replication of METR's study finds GPT-5 flags nearly 98% of reward hacking at a 5% false-alarm rate. At scale, that still means thousands of bogus warnings a month.

Nov 8, 2026 · 2 min