I had a similar problem when I also tested the eval board with a PowerProfiler Kit 2, have a look at my thread here:
It was primarily the Ethernet component that has a high power consumption during my testing, after desoldering It lowered the power consumption a lot, then you still have the PMIC and fuel gauge using power typically.
The Device OS version for the original B523 testing wasn’t recorded in the internal document, but it would have been old. However, The B404X was recently tested with Device OS 5, and the numbers with cellular off were identical to the B523 numbers, so I don’t think the Device OS version should matter too much.
As I am not able to verify the 55uA and it is paramount for the project I had in mind, I have decided not to go ahead with it.
It is great to see the increased focus on low power applications. It is cool that ultra low power mode retains the date and time. I hope in the future sleeping modes will become part of release/regression tests for LTS releases.
If you want to work with real low-power applications, I suggest to look at the Nordic Semi NRF9160 (down to 5 microampere in sleep!) and it has LTE-M and NB-IoT connectivity for low-power data transmission. It can be found in a prototype kit called Nordic Thingy:91 - I’m testing it out as well, as we cannot develop low-power variants with the current Particle hardware.
Sleeping on it, I found the leaking culprit by measuring consumption w.o. B523 mounted, to be the 47Kohm R9 SD card detect pull-up.
Disabling the card detect with a plastic insert, and with B523 mounted, consumption during ultra low power sleep was 37.9uA on average over a minute for both the 3.3V and VCC rails combined powered by 3.6V.
This dropped down during the first minutes, and settled around 36.4uA average.