I wonder how modern models fair on NovelQA and FLenQA (benchmarks that test ability to understand long context beyond needle in a haystack retrieval). The only such test on a reasoning model that I found was done on o3-mini-high (https://arxiv.org/abs/2504.21318), it suggests that reasoning noticeably improves FLenQA performance, but this test only explored context up to 3,000 tokens.