I know nothing about the first ARM, but ARM2 of Archimedes (anno 1987) was significantly faster than a MC68k (both at 8MHz), both much faster than a 6502 at one (typical) or two (in the beeb) MHz.
A BASIC interpreter using the ARM 1 or 2 might not have been literally faster than machine code on a 6502 (certainly not for some silly micro benchmarks), but, the stated goal, allowing high level programming where earlier assembly was required, certainly was met.
Whereas I suspect that I am nowhere near the only person on this page who once disassembled ROMs on a BBC Micro. We can state, in contrast, that there was no such self-modifying code. Again, BBC BASIC was in ROM.
Those lucky enough to have a copy of Jeremy Ruston's book after all of these years, or the retrocomputing enthusiasts who still have working Beebs, could even tell you exactly where in ROM the code was that fetched the next token for execution.
I never actually owned a copy of the book, and somewhat envy anyone who still has a copy; although to compensate I do have part of one of my own disassembly listings still, buried somewhere. (-:
It also had more (16 user registers vs, counting optimistically, 3) and larger (32 bit vs 8bit) registers, compared to the 8 bit registers of the 6502, and a 3-stage pipeline.
I expect that means the BASIC interpreter kept the current program position in a register, where the 6502 one used memory (likely in self modifying code), and could fetch the next token in a single cycle vs at least 5 or so for the 6502 version.
Having a faster CPU and more memory also may have meant they could be smarter in the way programs get stored.
I guess all that combined means there are programs were that goal can be met, for example programs computing a Mandelbrot image.
ARM2 added a hardware multiplier.