Build2 distinct publishers3 min readPublished
Google's new model trains on sparse station observations instead of six-hour-old physics output, which is what buys the 5 km surface grid. Wind aloft still comes out at 25 km, the same as WeatherNext 2.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
The 5 km figure comes from a different output head, not a downscaled grid. WeatherNext 3 trains directly on sparse weather station observations, and it predicts station-level sparse coordinates natively [7][8]. That is what the documentation means when it quotes 0.05 degrees for "station-calibrated surface variables" [3]. So a 5 km value near your site is only as good as the station network near your site, and Google's own limitations page says a 5 km grid remains a model representation rather than a promise about a specific address [15].
Then the arithmetic. WeatherNext 2 ran a 25 km grid in 6-hour increments [6]. Going from 0.25 to 0.05 degrees is five times finer per side, so twenty-five times the cells per unit area [17]. Hourly steps multiply that by six again, which puts one field for one member at roughly 150 times the values it used to carry [18]. Keep the full 64-member ensemble out to 15 days and a single variable is 360 timesteps across 64 members, about 23,040 member-steps per run [11][22]. That is what the resolution claim costs, and it lands in your object store and your joins.
The two published framings differ in a way worth reading closely. The documentation quotes a 2.5x to 5x improvement [5]. The blog post says roughly five times sharper [6]. Both are true of different variables. Temperature and moisture get 5 km, other surface fields 10 km, and atmospheric variables such as wind speed 25 km [4]. That last tier is the grid WeatherNext 2 already produced [19]. The gridded core at 0.1 degrees is 2.5x per side, so 6.25x the cells [20].
This is where the renewables framing needs a spec sheet. WeatherNext 3 outputs 100-metre wind speed, cloud cover and surface solar radiation downwards directly [13]. Neither source says which tier 100-metre wind sits on. If it is filed with atmospheric wind, a wind desk is getting WeatherNext 2's grid with a better ensemble bolted to it.
The strongest engineering here is upstream of the grid. Models trained on numerical weather prediction output inherit a six-hour data lag, which biases exactly the fast variables that matter, rain and surface temperature [9]. WeatherNext 3 ingests live one-hour geostationary satellite mosaics and emits a new forecast every hour against the most recent observations [8][10]. That is a mechanism, and it explains the precipitation result better than the score does.
About that score. Google reports up to 50% off Brier and CRPS, credited to training against ECMWF reanalysis, NASA IMERG and its own satellite-radar reanalysis [12]. The material does not state the baseline, the verification dataset, the region, or the lead time. For the number to transfer you would need held-out verification at gauge density comparable to whatever produced it, and the "most accurate" framing rests on live evaluations by Brightband, cited by Google itself [2].
A forecast that refreshes hourly and a planning system that refreshes nightly meet once a day, and the other twenty-three runs are decoration. The dev.to write-up makes the same point in plainer terms: the value comes from wiring a forecast to a concrete decision, such as when to dispatch a crew or alter a route, and a sharper number does not create that workflow [21].
Ranked by verification strength, evidence, and original report placement.
Google DeepMind and Google Research introduced WeatherNext 3, which Google's blog describes as the most advanced and accurate global weather model to date, according to independent live evaluations by Brightband.
According to Google's WeatherNext 3 benefits and limitations documentation, the model reaches resolution of up to 0.05 degrees, roughly 5 km, for station-calibrated surface variables, while core gridded fields reach 0.1 degrees, roughly 10 km.
Google's blog states that key surface variables such as temperature and moisture are available at 5 km resolution, other surface variables at 10 km, and atmospheric variables such as wind speed at 25 km.
Google says its training approach produces a 2.5x to 5x resolution improvement over WeatherNext 2, which operated at 0.25 degrees.
Google's blog says WeatherNext 3 provides a global weather picture roughly five times sharper than WeatherNext 2, which produced forecasts on a 25 km grid in 6-hour increments.
WeatherNext 3 provides hourly forecast timesteps up to 15 days ahead, with global coverage.
Distinct publishers with included, body-backed reporting in this cluster.
blog.google
1 article · September 3, 2026
dev.to
1 article · September 3, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
product
WeatherNext 3 moves Google from weather research to weather supplier1 distinct publisher
science
ECMWF's AI and physics ensembles swap places between five days out and two1 distinct publisher
build
Anthropic's protein run is checkable, which is rarer than the 26.8% hit rate2 distinct publishers
build
Thirty runs of one prompt put a 34-point error bar around a brand's mention rate1 distinct publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Precise numbers, single origin
Every figure in this story — the tier resolutions, the 15-day hourly horizon, the 64 members, the 50% precipitation gain — originates with Google, either in its announcement or in the benefits-and-limitations documentation that dev.to summarises. That is unusually specific for a launch, and the arithmetic checks out internally. What is missing is anyone outside Google: the Brightband evaluations that carry the 'most advanced and accurate' ranking appear as a name only, and the precipitation percentage has no comparison model or lead time attached to it.
Announced, not yet obtainable
There is a dated release and a sentence about forecasts becoming accessible across Google products worldwide, and that is the whole record. No product is named, no date given, and the documentation reviewed here sets out no pricing, access tiers or API path — which is why dev.to tells teams to treat the specifications and the access question as separate problems. Nobody outside Google is shown using this.
Sharper in the tier they photographed
The side-by-side UK temperature image is real and so is the 5 km surface grid, but 'roughly five times sharper' is the top of a 2.5x-to-5x range applied to the whole model, and the atmospheric tier sits at 25 km — precisely where WeatherNext 2 already was. A reader who cares about wind aloft has been sold a gain that does not exist for them. Against that, the architectural story is specific rather than decorative, and the dev.to piece supplies the deflation Google skips: a finer grid changes nothing until it reaches the decision it is supposed to inform.
The vendor and the vendor's integrator
Google is introducing its own model and grading it with a superlative; that is expected and legible. The subtler pull is at the other end, where the dev.to analysis closes by offering Scalevise's API and system-integration services. Its most valuable observation — that the payoff lives in the workflow, not the forecast — happens to be the argument for hiring the author. Both accounts profit if you conclude this model matters, and neither has any reason to publish the tier that stayed flat.
Firm on specs, thin on results
We can state what WeatherNext 3 emits, how often, at what resolution and in how many members with real confidence — those numbers are documented and consistent across both accounts. We cannot yet say how well it forecasts, what it costs, or who will run it, and the arithmetic-derived points about grid cells and data volume are only as good as the specifications they rest on.