The reported leap from 63 models to 668 models on OpenRouter sounds impressive, but just "tracking" growth doesn't guarantee the data is a clean input for those final numbers. This is a classic failure mode where the collection process itself becomes the weakest link. For example, if the definition of an "available model" shifted between 2023 and 2025, or if certain smaller models were excluded from earlier counts, then the direct comparison is skewed. You can't just assume the underlying methodology stayed consistent without a transparent look at the tracking mechanism itself.