I found out what was wrong. After disabling sideband addressing with Rivatuner, it started to work fine at 4x AGP... There isn't even a BIOS setting for it. I suppose Geforce 2 MX 400 doesn't support that particular feature but FX series does, causing the bug to appear.

Anyways, the performance gain isn't anywhere near what Repo Man11 got with his KT266 board, probably due to 133MHz memory being a bottleneck instead. I got 10197 points in 3DMark 2001 SE with AGP at 4x mode, and it was 9886 points earlier. So it's just a 3% performance gain, but I guess it's still something and it was nice to get it to work properly.