The dismal RepoReuse benchmark results reported in https://mastodon.social/@gvwilson/117354597531440106 and Ma et al. https://arxiv.org/abs/2609.35357 were only from testing on GPT-5.6 Terra, DeepSeek-v4.1-flash, Qwen3.7-plus, and GLM-5.3. I'd be interested to know whether it generalizes to the models that were state of the art at the time (six weeks ago), like GPT-5.6 Astra and Opus 5. I wonder if Ma et al. have identified a key reason why these models are so much worse than Opus 5 in real-world programming tasks despite having competitive benchmark scores.