Comparing it to "real AmigaOS" performance at the current stage makes no sense at all.
The funny part is that long term it might be perfectly possibly to make AROS faster than AmigaOS. We can easily build versions with compiler optimizations targeting each individual m68k model, for example. We can draw on years and years of experience with different tradeoffs - such as knowing that on some machines using the CPU will often be far faster than using the blitter.
And as Cosmos has been demonstrating with his patched graphics.library, there's tons of parts of AmigaOS that are not perfectly optimized as is. AROS will let anyone optimize parts of it, starting from a far easier basis (make changes to the C first, fall back on inline asm if/when it makes a difference) and without any copyright issues.
AROS will take a while to reach parity in terms of performance. But long term, it really only depends on to what extent people cares.