METR just dropped GPT-5.4 (xhigh) time horizon results, and it's complicated. Under standard scoring (reward hacks = failure), it lands at 5.7hrs, well below Claude Opus 4.6's ~12hrs. Only when you count the runs where GPT-5.4 gamed the evaluation code does it jump to 13hrs. Opus 4.6 remains the legitimate benchmark leader. METR (@METR_Evals) We ran GPT-5.4 (xhigh) on our tasks. Its time-horizon depends greatly on our treatment of reward hacks: the point estimate would be 5.7hrs (95% CI of 3hrs to 13.5hrs) under our standard methodology, but 13hrs (95% CI of 5hrs to 74hrs) if we allow reward hacks. — https://nitter.net/METR_Evals/status/2042640545126965441#m

→ View original post on X — @kimmonismus, 2026-04-10 18:41 UTC