Distros:
Redhat (2.6.18-128.2.1.el5 #1 SMP Wed Jul 8 11:54:47 EDT 2009 x86_64 x86_64 x86_64 GNU/Linux)
Ubuntu (2.6.24-24-386 #1 Tue Aug 18 16:24:26 UTC 2009 i686 GNU/Linux)

When I run 'top' in batch mode (specifying 1 iteration), the CPU % idle (%id) value is always the same and not accurate. Even when in interactive mode, the initial data display has incorrect CPU %id data. In interactive mode, the CPU %id value gets updated and by the third or fourth iteration, it looks to be accurate.

I installed the sysstat package on both systems. Using 'mpstat', the CPU utilization percentages don't match what 'top' displays in interactive mode. I can have two shells side-by-side, one with top in interactive mode and one running 'watch -n 3 mpstat'. They are both updating every 3 seconds, but the CPU utilization data is different for each.

Why is the first CPU utilization displayed by 'top' inaccurate?
Between 'top' and 'mpstat', which is most accurate/reliable?
Is there a better utility to fetch CPU utilization data?
Is there a reasonable way for me to calc my own CPU utilization data from /proc/stat?

Because a number doesn't look like what you want it to doesn't make it inaccurate. /proc/stat maintains numbers since last boot - by definition the first interval is ...
Sampled data on a non-idle machine (below 100%) is unlikely to be consistent.

I like sysstat due to the history data it affords - top also has its place, but is a bit "heavy", and tends to affect the numbers itself on lightly loaded machines.
Re-reading /proc/stat and determining the difference over the interval will allow you to write your own - same as everyone else does it.

Thanks for the info. I'm actually looking for real-time CPU utilization data. Do both 'top' and 'mpstat' report an average over time?

Yep, because that's the only thing that makes sense. What do you want, three billion reports a second on what each processor just did?

If you find existing tools insufficient, look into /proc/stat as syg00 suggested. That's the raw data, straight out of the kernel. It doesn't get more precise than that.