Huh? Running it backward as a thermoelectric generator is low efficiency (Carnot limit) but the efficiency pumping heat from e.g. freezing to room temperature isn’t that awful. Picking some completely random model from DigiKey, if the hot side was at 27c/80f, and you wanted the cold side at 0c, 2.4 amps at about 7V (16.8W input power) would let you remove about 15W from the cold side with a deltaT of 30c (i.e. the cold side at -3c/27f) per the graphs on page 5. That’s a COP of about 0.9 for cooling. You’d need to remove about 30W from the hot side.

In practice, cooling the hot side is difficult so you might instead use the Th=50c graph, at which point 2.4A @ 8v (19.2W input power) would give 3-4W cooling power with the cold side at 0c, for a COP of about 0.18. But nowhere near your numbers.

If you put some decent thermal mass on the hot side and operated intermittently, as long as you drained the condensation, you could get closer to the Th=27c performance.

You're thinking of carnot efficiency power cycles where you are removing energy from a fluid, maybe? I meant exactly what I said with their COP.

Refrigeration/heat pump cycle COP's can be all over the place. Often they are much higher than 1.0. Where a refrigeration cycle's COP is calculated as Q(cold) / Work(electrical), many heat pumps can move more watts than they take to run. TEC's are no exception.

TEC's in particular are especially sensitive to the total temperature differential between hot and cold plates when it comes to COP. They hate making temperature differentials much higher than 30-40 kelvin. Most efficient when the temperature delta is like 10K or less. COP 0.6 is about typical for a TEC running across the fairly large temperature differential that you would need to get air down to dew point and condense water, at a high current required to make the TEC's cost effective. Example graph of COP vs temperature differential