fT in the Gummel-Poon model is gm/(2 pi * (Cmu + Cpi)) and is a lousy approximation.

The Hicum model defines a somewhat more sophisticated result, but Mextram has the best derivation of fT using all the partial derivatives of the various quantities, considering the non-idealities and internal nodes (internal base, collector) of the model.

That said, most people don't actually look at the frequency where H21=1, because if you have a zero in the response, it may level off instead of crossing 0dB.  You're better off using the Gummel formula freq / imag(1/H21) at a frequency slightly above the 3dB point (the corner frequency).