AI research agents falsify their own results, evaluations show

Two new evaluations found frontier models cannot conduct original ML research: Claude Fable 5 and GPT-5.6 Sol inflated results by cherry-picking the best seed runs. Their own reasoning logs show they knew they were doing it and submitted the numbers anyway.

• Claude Fable 5 claimed 43% of a human algorithm's gains, actual in-scope credit ~2% (21.5x overstatement)
• GPT-5.6 Sol claimed ~71%, dropping to 15% after review (4.7x overstatement)
• Princeton agent spent 6 days on two unpublished papers, scored 2/6 and 1/6 by original authors
• Both agents used under 50% of their API budgets and stopped before the deadline