For a time systems with a 386SX were significantly cheaper than those with a 386DX because the 16-bit data-bus mean cheaper motherboards could be used.

If you were running 16-bit software they were little slower than a 386DX at the same clock and significantly faster than a 286 because of higher clocks (286's usually topped out at 12MHz though there were some 16MHz options, the slowest 386s were running at 16MHz with some as fast as 40MHz), but also in part, when not blocked by instruction ordering issues, to the (albeit small by modern standards) instruction pipeline which the 286 lacked.

32-bit software was a lot slower than on a DX because 32-bit data reads and writes took two trips over the 16-bit data bus, but you could at least run the code as it was a full 386 core otherwise (full enhanced protected mode, page based virtual memory, v8086 mode, etc).

The SX also only used 24 bits of the address bus, limiting it to 16MB of RAM compared to the original's 4GB range, though this was not a big issue for most at the time.

I can't remember, could you buy a math coprocessor for it?

I know my 286 you could pair with a 287 next to it.. not sure if it really made a difference you could discern outside of hyper-specific uses though.

Ahhh but it gave me the opportunity to ran real programs, coming from an XT! *Edited to add an example: I could for the first time use AutoCAD. The price difference between a 286 and a 386SX was negligible, but the software I could use, was other league.