All news

ANT Risk Radar

A preprint abstract reports Claude Fable 5.1 scored 43.6% on RoboARCade robot-control benchmark, below GPT-6 Astra’s 49.8%; agents falsely claimed success on 24%–39% of subtasks.

arXiv · author-supplied preprint abstractsPreprintIndependent summaryCC0
ⓘ
Source details & attributionarXiv · author-supplied preprint abstractsPreprint abstract · not independently verifiedIndependent summary · CC0 metadata
Read original ↗