Tabular foundation models (TFMs) produce predictions from a labelled context in a single forward pass. A recent benchmark claimed they decisively outperform tree ensembles for customer churn, but it compared them with untuned gradient‑boosted trees (GBDTs) and reported only ranking metrics, whereas a retention campaign is judged by monetary profit. We re‑evaluate TabPFN‑2.5, TabPFN‑3.5, TabPFN‑3.5‑Fast and TabICL v2 against Optuna‑tuned GBDTs on five public telecom churn datasets using ten seeded splits. Three questions are addressed: does a model rank churners well, does that ranking yield higher expected maximum profit (EMP), and can its probabilities drive a threshold policy? With a split‑aware test and Holm correction over 20 pre‑declared comparisons, TabPFN‑3.5 achieves significantly higher EMP than the best tuned GBDT on Cell2Cell (+1.9%) and Iranian (+7.6%), and these gains hold against each of the three tuned GBDTs separately. On Orange it leads in all ten splits (+14.3%) but the gain is not significant after correction (Holm p = 0.155). The splits resample the same datasets, so the results do not generalise to independent operators. TabICL v2 beats the best tuned GBDT only on Iranian, and underpredicts Orange churn by 2.1 points on a 7.3% base rate. A substantially larger tuning budget and a duplicate‑aware split leave the main conclusions unchanged. Retuning the baseline reduces the earlier headline gap from 6.1 to 4.0 PR‑AUC points without closing it. The licence of the TabPFN‑3.5 weights we used (License v1.0), not scoring cost, is the only barrier between its gains and production deployment.
Review: The paper provides a rigorous profit‑oriented reassessment of recent tabular foundation models, highlighting both their potential advantages and the importance of proper baseline tuning and statistical correction.