Tabular foundation models are rapidly superseding gradient-boosted decision trees as the dominant approach for structured data prediction, a shift confirmed by an independent audit of the TabArena benchmark. Led by TabICLv2, a lightweight model from Inria's SODA team, zero-shot transformer architectures now claim the top positions on community leaderboards, outperforming heavily tuned traditional baselines with significantly lower inference costs. TabICLv2 operates as an in-context learner, a single pretrained transformer with approximately 28 million parameters that predicts target columns without gradient-based training. The model ingests labeled rows as context and attends to them during a forward pass, effectively functioning as a learned k-nearest-neighbors mechanism. Pretrained exclusively on synthetic data generated via structural causal models, TabICLv2 avoids benchmark contamination. Its open weights, released under the BSD-3 license, enable full reproducibility. In an independent evaluation covering the full TabArena Lite protocol, the auditor reproduced TabICLv2's performance on a single AWS A10G GPU for approximately $2. The model achieved an Elo rating of 1559 against the official 1575, falling well within bootstrap confidence intervals and confirming the integrity of the leaderboard claims. Across 51 datasets, TabICLv2 won 40 comparisons against the best tuned and ensembled gradient-boosted models, including LightGBM and CatBoost. The foundation model demonstrated superior cost-efficiency, requiring seconds to adapt to new data versus hours for tree-based hyperparameter search, while maintaining median prediction times comparable to optimized CPU implementations. The benchmark landscape reveals a distinct paradigm inversion. As of the 2026-07-15 snapshot, every single-model entry exceeding the performance of tuned gradient-boosted trees is a tabular foundation model. Google Research's TabFM leads the board but lacks technical documentation and carries a non-commercial license. Prior Labs' TabPFN series offers strong performance but restricts commercial usage. TabICLv2 stands as the strongest contender with unrestricted open access. Performance regimes clearly delineate the strengths of each architecture. Tabular foundation models excel on tables with fewer than 100 features and moderate row counts. However, gradient-boosted trees retain advantages in high-dimensional spaces exceeding 100 features and datasets with high-cardinality categorical variables. Audits indicate that ensemble routers can improve accuracy by selecting between complementary foundation models, though validation-based selection fails to optimize routing between foundation models and tree-based baselines. The rise of these models challenges the default workflows of the past decade. While tabular LLMs in the table-native category, distinct from prompting large language models with serialized text, demonstrate clear advantages, practitioners must weigh the regime-specific limitations. For most standard tabular tasks, zero-shot foundation models now offer a faster, more accurate, and fully open alternative to bespoke tree tuning.