WeatherNext 3: Reportedly the First Hourly 5km Global AI Forecasts

Global weather models have historically faced a trade-off between how finely they slice the map and how often they can afford to re-run. Coarser grids and infrequent updates are cheaper to simulate; finer grids and hourly refreshes are what forecasters actually want for anything that moves fast — a flash flood, a sudden wind shift, a storm cell that spins up in an afternoon. Reports circulating around September 6-7, 2026 describe Google DeepMind's next release in its WeatherNext family, WeatherNext 3, as reportedly the first AI model to deliver global forecasts at roughly 5km resolution, updated hourly. We have not seen a primary DeepMind blog post or paper confirming these exact figures, so this piece treats them as reported rather than officially verified — but the shape of the claim fits squarely into the pattern explainx.ai tracked in DeepMind's WeatherNext Cyclones release just one month earlier.

TL;DR

What "5km, hourly, global" actually means

Put the numbers next to what came before. WeatherNext Cyclones, the model explainx.ai covered in August 2026, runs on 28x28km resolution inputs — already surprisingly coarse for its accuracy, a result DeepMind itself called an open research question. Global numerical weather prediction (NWP) systems that meteorological agencies have relied on for decades typically operate at resolutions in the tens of kilometers, with full model runs a handful of times per day rather than hourly, because simulating the physics of the atmosphere at finer grids and higher frequency multiplies compute cost quickly.

If the reports about WeatherNext 3 hold up, a jump to 5km resolution is roughly a 30x finer grid than WeatherNext Cyclones' inputs, and hourly updates are a step-change in cadence compared to the few-times-daily rhythm of traditional global models. In plain terms: a coarser grid effectively averages weather over a large area, so a fast-forming storm cell, a flash flood over a single watershed, or a localized wind shift that changes wildfire behavior can fall inside one grid square and get smoothed away or reported hours late. A finer grid updated every hour, in principle, can track these fast, local phenomena much closer to real time — the exact category of event that coarser, less frequent forecasts have historically struggled with.

That said, resolution and accuracy are not the same thing, and it's worth being honest about what isn't confirmed here: the exact forecast horizon WeatherNext 3 covers, what baseline it was benchmarked against, and what compute cost is involved in running it hourly at global scale are all details we don't have from a primary source as of this writing. Readers should treat "5km hourly global" as a reported capability claim, not a validated benchmark result, until DeepMind publishes its own numbers — the same caution warranted for any pre-announcement report in this space.

Why this is an AI-vs-traditional-methods story, not just a weather story

The reason "hourly global at 5km" is even conceivable is the same reason GraphCast mattered in 2023 and WeatherNext Cyclones mattered in August 2026: AI weather models don't simulate atmospheric physics directly the way traditional NWP does. They learn statistical patterns from historical atmospheric data during training, then produce a forecast in a single, fast inference pass. WeatherNext Cyclones' full 1,000-member, 15-day ensemble reportedly ran in under a minute on a single TPU — a fraction of the compute and wall-clock time a physics-based ensemble at that scale would require.

That pattern — dramatically more capability per unit of compute once a model is trained — is the same thread explainx.ai has tracked across AI's role in climate and energy sustainability and the compute-cost math behind model selection. Reportedly extending that efficiency to hourly, 5km global coverage would be a meaningful demonstration that the "AI forecast at a fraction of the cost" story generalizes beyond cyclones to weather forecasting broadly, not just one high-value storm category.

How WeatherNext 3 fits the WeatherNext lineage

• GraphCast (2023) — the original breakthrough: 10-day global forecasts in under a minute on a single TPU v4, matching or beating operational NWP beyond day 7.
• GenCast (2024) — added probabilistic ensemble forecasting to quantify uncertainty and extreme-weather risk.
• WeatherNext Cyclones (August 2026) — unified track and intensity prediction for tropical cyclones in one model, running on 28x28km inputs, open sourced with a Colab-runnable mini version. Read explainx.ai's full breakdown of that release.
• WeatherNext 3 (reportedly September 2026) — the same underlying research direction pushed toward general-purpose global coverage at far finer resolution and hourly cadence, rather than a cyclone-specific model.

This also sits alongside Google Research's Planetary Prediction Engine, a separate Earth AI project under the same umbrella that turns natural-language queries into custom geospatial regressions rather than producing weather-field forecasts directly — a reminder that "Google Earth AI" now spans several distinct model families rather than one product.

What builders in weather-adjacent domains should watch for

Higher-resolution, hourly forecasts are exactly the kind of input that agriculture tech, insurance and climate-risk tooling, logistics and routing software, and disaster-response systems would want to build on — but there's a real gap between "a model exists" and "a model is accessible." WeatherNext Cyclones and WeatherNext 2 were open sourced with downloadable weights, and DeepMind has separately exposed forecasts through its public Weather Lab interface without requiring anyone to run the model themselves.

Whether WeatherNext 3 follows that same open-access playbook, ships behind a hosted API, or stays a research-only release hasn't been confirmed in what's circulating as of this writing. If you're evaluating whether to build a product around it, the honest posture right now is: worth watching for an official access announcement, not something to assume is already integrable. Teams in AI + climate tech building on top of weather data specifically should keep an eye on DeepMind's own channels rather than a third-party API wrapper appearing prematurely.

What people are asking

Is WeatherNext 3 a replacement for WeatherNext Cyclones? No — they appear to solve different problems. Cyclones is purpose-built for tropical storm track, intensity, and wind structure; WeatherNext 3 is reportedly a general global forecasting model. An agency tracking an active hurricane would likely still want the cyclone-specific model's specialized ensemble output rather than a general-purpose forecast.

Does "first hourly 5km global AI forecast" mean it beats every existing model? Not necessarily, and that claim isn't the same as an accuracy claim. Resolution and update frequency describe what the model outputs and how often — they don't by themselves prove the forecast is more accurate than a coarser model's output at the same lead time. DeepMind's own WeatherNext Cyclones paper is a useful caution here: it found a coarser grid outperformed prior fine-resolution approaches for cyclone intensity, which is a reminder that resolution numbers alone don't settle an accuracy question.

Where can I verify these numbers myself? Search for an official Google DeepMind blog post or paper under the WeatherNext or Google Earth AI name — at the time of writing we have not located one, and this post is explicitly flagging the reported figures as unverified pending that primary source. Google DeepMind's site (deepmind.google) is the place to check for an eventual official announcement.

Does this replace human forecasters or official weather warnings? No — consistent with DeepMind's stated position on WeatherNext Cyclones, these models are described as tools that support a "collaborative weather forecasting ecosystem" alongside human meteorologists at national agencies, not a replacement for official warnings. Always defer to your local meteorological agency for actual storm and weather alerts.

Related reading

• DeepMind's WeatherNext Cyclones gets a full extra day — the direct predecessor covered here, including its 28x28km resolution finding and open-source release
• Google PPE: a geospatial AutoML agent — the other major Google Earth AI release from August 2026
• Top 10 AI + climate tech startups to watch in 2026 — where climate-risk and weather-data tooling fits into the broader startup landscape
• AI and climate change: the paradox of a tool that causes and fights the crisis — GraphCast's origins and the compute trade-offs of AI-for-climate work
• The model-selection energy math — why coarser, faster inference matters beyond weather specifically
• Google TimesFM-3: multivariate time series forecasting — another September 2026 Google forecasting-model release, for general time-series rather than atmospheric data

This post describes WeatherNext 3 based on reports circulating around September 6-7, 2026; we have not located a primary Google DeepMind announcement confirming the exact resolution, forecast horizon, or access model, and specifics here should be treated as reported pending official confirmation. For official storm and weather warnings, always defer to your local meteorological agency or national weather service.