The man page
... values between 0 and 1023. This yields a resolution between readings of: 5 volts / 1024 units ...
... which supports what mart says.
But beware the fence-post error
here: it may well be 1024 values, but it's still only 1023 spaces.
For simplicity let's say the values were only 0, 1 and 2. That's 3 values clearly (0, 1, 2), but 3 values have only 2 gaps: the 0-1 gap and the 1-2 gap.
So if 0V is 0 on my pretend 0-1-2 analog sensor scale and 5V is 2, then it's clear that it's 2.5V per gap... position 0 is 0V, position 1 is 2.5V and position 2 is 5V. So the 5V is spread over 2 gaps not 3 positions.
By that logic, where the gaps are what count not the number of values (counting the start and end fence-posts), the 5V is divided over 1023 gaps not 1024 positions.
Edit: my calculator tells me 5/1023 is 0.0048875855327468 and 5/1024 is 0.0048828125, so I'd call those 0.00489 and 0.00488 both of which are 0.0049 so who cares in practice? But yes- it's good to be clear on the theory....