Eddy-correlation fluxes are compared to air-sea fluxes predicted by a widely used bulk flux formulation without wave-state effects. Systematic discrepancies are found. For example, the model approximately equates the roughness lengths for heat and moisture; however, the observed roughness length for heat (zoh) exceeds that for moisture (zoq) by an order of magnitude or more, except in the strongest wind-speed conditions. This is apparently due to the dynamic nature of temperature, which dominates buoyancy generation of turbulence in these data sets. The observed correlation between temperature and vertical velocity fluctuations generally exceeds that for moisture. For 10-m wind speeds above a threshold value of 12 m s-1, zoq exceeds zoh apparently owing to enhanced moisture flux associated with the onset of wave breaking coupled with advection of cold dry air from land. In near-collapsed turbulence, the observed momentum flux is smaller than predicted, and there is no clear indication of a smooth flow viscous regime. The scatter between observed and bulk fluxes generally decreases with averaging the observed fluxes over greater length scales even with variations in sea-surface temperature (SST). The reduction in random flux sampling errors more than compensates for capturing increased surface heterogeneity with increasing averaging scale. Since similarity theory does not apply to heterogeneous surfaces, the bulk model does break down in the extreme case where the averaging window includes a sharp SST front. The response of the flow to changes in SST is presented for different amplitudes of SST variability. The change in vertical structure and acceleration of the low-level wind over warm pools is discussed.