WeatherNext 3 Accuracy: Benchmarks, Limits and How to Check It
Google DeepMind reports that WeatherNext 3 delivers up to 50% more accurate precipitation forecasts a day or more ahead and up to 60% better probabilistic precipitation scores against the IMERG reference at early lead times. Those are specific, checkable claims, but they describe named baselines rather than a promise for every town. This independent guide explains what the numbers measure, where they come from and the limits they carry, then shows how to verify a forecast where you live. If you want the model input first, start with what changed in the WeatherNext 3 model or how to read the hourly forecast.
7-day free trial, then $19/mo · cancel anytime · compare plans
The accuracy numbers Google publishes
Drawn from Google DeepMind's launch announcement and the WeatherNext 3 paper. Each figure is measured against a named baseline, and the baseline matters as much as the number.
50% Reported improvement in precipitation forecast accuracy a day or more ahead, with the largest gains in regions where forecasts were historically least reliable.
Reported improvement in precipitation forecast accuracy a day or more ahead, with the largest gains in regions where forecasts were historically least reliable.
60% Reported CRPS improvement against NASA IMERG in medium-range evaluation. Google also reports 30% against MRMS radar and 10% against rain gauge measurements at early lead times - all Google's own evaluation.
Reported CRPS improvement against NASA IMERG in medium-range evaluation. Google also reports 30% against MRMS radar and 10% against rain gauge measurements at early lead times - all Google's own evaluation.
5 km Station-targeted temperature and humidity now resolve at 5 km, other surface variables at 10 km, against a 25 km grid for WeatherNext 2. Sharper grids matter most near coasts, valleys and mountains.
Station-targeted temperature and humidity now resolve at 5 km, other surface variables at 10 km, against a 25 km grid for WeatherNext 2. Sharper grids matter most near coasts, valleys and mountains.
What an accuracy claim actually measures
Five questions separate a meaningful benchmark from a headline. Ask them of WeatherNext 3 or any other model.
The metric
CRPS scores a probabilistic forecast: it rewards confidence that turns out to be justified and punishes overconfidence, across many locations and times rather than one dramatic case.
The baseline
A 60% improvement means versus a specific reference - IMERG satellite estimates, MRMS radar or rain gauges. Every baseline has coverage gaps, so any single comparison stays partial.
The region
The biggest reported gains are where forecasts have historically been weakest, including parts of Latin America, Africa and Asia-Pacific. Your local baseline may be stronger or weaker than the global story.
The lead time
Improvements vary by how far ahead you look. The 10% rain gauge figure applies to early lead times; day-ahead and multi-day windows should be judged separately.
The evaluator
Alongside the paper, live leaderboards such as Brightband's Open Weather Bench score operational runs against observations continuously, which is where you can watch the claims age in public.
How to sanity-check WeatherNext 3 where you live
No benchmark tells you whether the forecast matched the ground under your feet. These comparisons can.
Pick a nearby reference
Find a weather station within a few tens of kilometers of your location. Close is not identical: coastlines, valleys and cities each shift rain and temperature at the 5 km scale.
Track one rainy day hour by hour
Compare when rain starts and stops against the hourly forecast from that morning. Timing and boundaries are exactly where AI precipitation forecasts improved most.
Compare like with like
Check the variable the forecast reported for that spot, not a nearby town's reading. Compare temperature to temperature, precipitation to precipitation.
Look across several days
One miss proves little in either direction. Ten comparisons across different weather patterns tell you far more about local reliability than a single event.
Limits worth knowing before you rely on it
WeatherNext 3 is a major step, not a crystal ball. These caveats are the normal territory of forecasting, and the model's authors acknowledge them too.
Safety decisions belong to official agencies
Google's launch note directs people to local meteorological agencies for warnings and public safety advisories. Treat model output as guidance, never as the warning itself.
Verification is harder where data is sparse
Gauges and radar cover land unevenly, and the open ocean cannot be checked the same way. Reported improvements rest on the best available references, not uniform global truth.
Not every variable improved equally
The sharpest gains are in precipitation and station-level fields. Wind at 10 km and upper-air layers around 25 km are coarser, and local extremes remain the hardest case for any model.
The figures move as evaluation continues
Accuracy statements are snapshots of a live process. The paper, the leaderboards and Google's developer guide are the places to re-check before quoting numbers.
WeatherNext 3 accuracy questions
Short, sourced answers about how accurate the model is and how to judge it.
Still comparing models? See what changed in the WeatherNext 3 model or learn to read the hourly output.
Test the forecast where you live
A benchmark is a starting point. Start the 7-day free trial, watch a week of hourly forecasts for your own location, and judge the accuracy against what actually happens. For background, read the model page or how to read the hourly forecast.
Related guides
The pages that go with this one, chosen by what each one answers.
