Lets have the 2-tone input power reduce 1dB. For IM3, the shoulders (F1+-2*F2, 2*F1+-F2) will drop 3dB. The difference drops 2dB. Lets call this difference "2X"
You've got the heads at -105, the shoulders at -128; difference is 23dB, which is our "2X". Thus "X" is 11.5dB.
To compute IP3, add X to the input power(the heads), thus IP3 is -105 + 11.5 = -93.5dB.
To make this work, the 2-tone inputs are set to the SAME power levels, and combined in stripline combiners or resistive summers. Since Frequency Generators will have some intolerance for unexpected frequencies imposed on their outputs, use some small attenuators between the 2 FreqGens and the combiners. Unless you are testing at -105dBm, as you are here.
The purpose of using TWO STRONG TONES is to examine how the energy in the shoulders might become strong enough to cause bit errors in the channels allocated to the shoulders [F1+-2*F2, or 2*F1+-F2]. Thus 1,000MHz and 999MHz, given some IM3, produces 998MHz (2*999 - 1,000) and 1,001MHz (2*1,000 - 999).
If the 2 tones have equal power, the intermods will strongest.
Only one set of measurements (powers of the 4 peaks) is needed, as shown in the example above with -105 and -128. Additional measurements [again, the 4 peaks], with "2X" computed, cut in half, and added to the input 2-tone powers, will give more confidence.
To catch localized VSWR peaks/dips, try several combinations of head-frequencies across the band.