Ive noticed this zoom dependence as well. At btschumy mentions its troubling because it unclear which value to trust, especially when there is a large discrepancy. I've seen over a 3x variation in forecasted PM 2.5 where it means the difference between Moderate and straight up Unhealthy air quality. Can this be interpreted as meaning the uncertainty in the forecast is even greater than this 3x discrepancy? This seems reasonable for a forecast days out but I've seen this discrepancy for the current time.
btschumy has a good point on grib file resolution possibly being sent at a lower resolution when zoomed further out. Can you confirm this is the case? Otherwise, I'm struggling to understand why the interpolated value would vary with zoom level. The only other explanation I can think of is if the coordinates of the chosen point are rounded more as you zoom out. For example, at a close zoom, say the interpolation is calculated by rounding the coordinate to the nearest thousandth of a degree. But zoomed further out the interpolation is calculated by rounding to the nearest tenth of a degree. In this case the computed forecast location could vary by a few miles which could make a huge difference.