It is very clear in that chat logs (which include reasoning traces) that the model knew that, knew what the last election it knew about was, and answered correctly based on its cut off initially. Under pressure to answer about an election that was not within its knowledge window it then confabulated a Biden 2024 victory, which it dug in on after being contradicted with a claim that, based on the truth at the time of its knowledge cutoff, was unambiguously false ("Joe Biden did not run") He, in fact, did run for reelection, but withdrew after having secured enough delegates to win the nomination by a wide margin on July 21. Confabulation (called "hallucination" in AI circles, but it is more like human confabulation than hallucination) when pressed for answers on questions for which it lacks grounding remains an unsolved AI problem.
Unsolved, but mitigated by providing it grounding independent of its knowledge cutoff, e.g., by tools like web browsing (which GPT-OSS is specifically trained for, but that training does no good if its not hooked into a framework which provides it the tools.)