Keep in mind that programs like G*Power provide an estimate of the required sample size. The estimate is only as good as the values that you put into the program. So Bob ran an experiment in 1973 with five replicates. I now use Bob's results for mean and standard deviation to enter into G*Power. The first question you should ask is how accurate is an estimate of the mean and standard deviation based on a sample size of five?

The estimate is a lower bound. It will seldom be THE right answer.

When in doubt, go for more samples. It is easier to argue that a statistically significant effect is too small to be relevant than it is to argue that "this" important treatment effect would have been significant had I taken more samples.

No matter what G*Power states, always look at the literature. If G*Power indicates that you need a sample size of 8, but all the equivalent literature reports experiments with sample sizes ranging from 12 to 75, then you will be in trouble using a sample size of 8. Be careful using a smaller sample size if G*Power indicates a sample size of 200 but the literature reports experiments that have sample sizes of 12 to 75.

With your experiment there are three potential outcomes.

1) Your results agree with the published literature.

2) Your results do not refute the published literature, but they also do not support it.

3) Your results contradict published findings.

In case #1, you are fine so long as you use a sample size that is no smaller than the smallest sample size in the most recent decade of published literature.

In case #2, it will be difficult to publish, though more journals are interested in inconclusive results so long as the methods are good. If you have average or above average replication, this might work.

In case #3, you need some reason to convince others that you did not make a mistake. This will be very difficult if your sample size is small because it is easy to reject a manuscript with a substandard sample size.

A part of planning sample sizes is risk management. What is the chance that you will discover something new that will contradict current paradigms? What risk are you willing to take that the faculty member next door will run a similar study with twice your replication and find contradictory outcomes? What risk are you willing to take to have inconclusive results? All of these issues are mitigated by having a larger sample sizes. Obviously this must be balanced with the cost and time involved with gathering more data.