The lunchtime umbrella check
Imagine checking a weather app at 8 a.m. to decide whether to carry an umbrella home. At noon, you check again for the same 6 p.m. rain. This is a hypothetical scene: the refreshed screen could show the morning prediction again, or a prediction recalculated with observations received since then.
Starting a fresh forecast
Google publicly announced WeatherNext 3 on September 3, 2026. The change to watch is that it starts a fresh forecast each hour.
Two clocks in one forecast
A forecast has two clocks: its starting reference time and the future time it describes. An 8 a.m. forecast might contain values for 5, 6 and 7 p.m. A fresh noon forecast could describe that same 6 p.m. Closely spaced hours on a chart alone cannot tell you whether the inputs have been refreshed.
Predicting the same 6 p.m. twice
What is the new reference time?
From an 8 a.m. forecast to a fresh noon forecast
Refresh inputs and start a new predictionWhich hour is being predicted?
5, 6 and 7 p.m. within each forecast
Compare 6 p.m. across different startsHow fresh observations enter
The research report describes recent satellite observations supplementing conventional atmospheric analysis: an estimate of the atmosphere combining observations with a physics-based model. More frequent satellite arrivals support hourly forecast starts. The satellites do not replace the analysis, and collecting and delivering observations still takes time.
What does 5 km describe?
The spatial scale also needs a label. In the documentation, approximately 5 km output applies to station-trained temperature and dewpoint predictions. Precipitation output is approximately 10 km. Reading 5 km as the resolution of every rain forecast would expand the claim beyond its scope.
A new question about the same hour
Return to the umbrella decision. The target remains 6 p.m., while the prediction’s starting point can change. An hourly forecast invites two questions: “Which hour is this about?” and “How recent are the observations used to recalculate it?” For the connection, follow the research report below to section 2.2, on observational input and hourly initialization.
This explains the structure from the announcement, research report and developer documentation. We have not tested a local app display or the accuracy of an umbrella decision.