I found these sensitivity measurements using a signal generator with 400 Hz tone and 30 % modulation:
200 kHz: -103 dBm
300 kHz: -102 dBm
600 kHz: -106 dBm
1000 kHz: -107 dBm
1500 kHz: -107 dBm
5000 kHz: -107 dBm
10000 kHz: -107 dBm
20000 kHz: -106 dBm
In terms of sheer numbers, the Excalibur is the most sensitive SDR I have measured, with a good 1uV sensitivity throughout most of the MW and SW spectrum.
So why does the S-meter read much lower than the Perseus? It is apparently not calibrated. When I fed a -107 dBm signal from the signal generator to the Excalibur, the S-meter read -115 dBm. This is roughly consistent with the difference between the Perseus and Excalibur S-meters.
Update July 20: Martin Kent at Winradio suggested that there might be a bug in version 1.06 that caused the erroneous signal level readout, and asked me to check the signal level with version 1.10. Indeed, the difference between the signal generator at Excalibur was reduced to 3 dB. My signal generator may not be 100 % calibrated, so I have no way of saying if the signal generator or the Excalibur was correct. Or any of them.
Now, for those who want to compare Perseus and Excalibur recordings, I have uploaded two 1 MHz recordings with 6120 kHz as center frequency. The recordings contain a lot of static due to nearby thundershowers, giving their AGCs something to work on. The signal levels aren't too high, with the exception of semi-local Murmansk 5930. This should be a good opportunity to compare selectivity.
(Update Aug 6: The download files are deleted due to server limitations. If you want to check out an Excalibur MW file, please download from here)