WeatherNext 3 Is Operational—but the Managed Inference Page Still Specifies WeatherNext 2

The decision is consume, pilot, wait or reject

Consume WeatherNext 3 operational forecast data when a global probabilistic feed can improve a bounded downstream decision and you can validate the exact variables, issue times and ensemble statistics. Pilot it against an incumbent on archived decisions. Wait if the requirement is on-demand custom WeatherNext 3 inference: Google’s current managed specification explicitly names WeatherNext 2. Reject any design that treats experimental output as an official severe-weather warning or allows a missing cycle to fail open.

This is AccessAllGPT guidance, not Google’s. “WeatherNext is on Cloud” hides at least two contracts: reading provider-generated WeatherNext 3 arrays and invoking a managed model to generate tailored forecasts. Architecture, pricing, quotas, data lineage, reproducibility and rollback differ between them.

What changed on September 3

Google launched WeatherNext 3 on September 3, 2026 and integrated forecasts into Search, Gemini, Maps, Google Maps Platform and Cloud. The developer overview describes a 15-day global probabilistic forecast, 64 ensemble members and hourly initialization. The paper dates the system after WeatherNext 2 and WeatherNext-Cyclones and says the production model was trained through June 30, 2026.

WeatherNext 2 used a 0.25° grid and six-hour cadence. WeatherNext 3 adds hourly 0.1° single-level output, 0.05° station-head queries in the evaluation, direct satellite input, precipitation heads and tabular cyclone outputs. The chronology matters: this is a new operational model family member, not merely a new visualization over the old feed.

The original finding is an access-contract split

Google’s overview says operational WeatherNext 3 data is available through Cloud Storage in Zarr, Earth Engine and BigQuery. On the same site, the model-specification page states: “The model that is accessible on Gemini Enterprise Agent Platform is WeatherNext 2.” Its table describes 0.25° resolution and six-hour initialization—the older contract.

That does not make WeatherNext 3 unavailable. It makes availability surface-specific. A team that needs provider-issued forecast arrays can proceed to a data-feed pilot. A team that needs custom initial conditions, ensemble sizing or on-demand generation must verify which checkpoint the managed endpoint actually serves and must not infer WeatherNext 3 from the umbrella product name.

How the model combines analysis and observations

WeatherNext 3 receives two analysis frames six hours apart and recent geostationary satellite mosaics. At 14 UTC in the paper’s example, the latest mosaic is 13 UTC while the latest analysis-derived input is valid at 08 UTC. The satellite stream is therefore roughly five hours fresher and lets the system issue a new forecast every hour.

The model is an encode–process–decode Functional Generative Network. Separate encoders map 0.1° and 0.25° modalities to a shared icosahedral processor mesh. WeatherNext 3 raises latent width from 768 to 1,024 and transformer depth from 24 to 32 layers, then generates 64 ensemble members with noise and epistemic dropout.

Hourly does not mean every internal field is hourly

The outer autoregressive step remains six hours. Single-level variables and selected pressure-level fields are predicted at hourly offsets inside that window, while other atmospheric variables retain six-hour steps. The continuous station decoder can be queried in space and time, but its outputs are target-only rather than fed back autoregressively.

This distinction predicts one documented failure mode: station-head jumps at the boundary between six-hour windows. Procurement language should name each required variable’s cadence and head, not advertise the whole product as uniformly hourly.

The evaluation is broad, but it is not one clean benchmark

The authors use full-year 2024 evaluation with a model trained through 2023 for most results, then a July 1–August 11, 2026 quasi-real-time comparison with the production checkpoint. Baselines change by task: WeatherNext 2, ECMWF ENS and AIFS ENS v2 appear in different slices. Ground truth includes HRES analysis, held-out stations, IMERG, MRMS, rain gauges and IBTrACS.

That breadth is useful, but it prevents one score from summarizing the release. IMERG is a target during training and therefore is not independent for the IMERG head; station coverage is geographically uneven; lower-resolution baselines are interpolated; and the six-week 2026 sample is explicitly too small for over-interpreting minor differences.

The strongest claimed gains belong to bounded slices

The paper reports roughly 5% average improvement over WeatherNext 2 on upper-level medium-range variables, interpreted as about six hours of lead-time gain at matched skill. For short-lead station temperature, it reports up to 30% lower CRPS than WeatherNext 2 and 40% than ENS. Its PARDIG precipitation head reports reductions up to 60% against IMERG, 30% against MRMS and 10% against rain gauges at early leads.

These are author-reported, metric- and dataset-specific results. The precipitation evaluation caps accumulation thresholds at 4 mm per six hours and leaves extreme precipitation to future work. “Up to” values cannot be used as an expected production uplift without preserving variable, horizon, region, metric and verification target.

Fresh satellite input buys hours, not certainty

In a latency-adjusted comparison against an otherwise identical six-hour initialization schedule, the paper reports a two-to-three-hour skill lead for early precipitation forecasts. That matches the average freshness advantage of hourly cycles and is operationally plausible for developing storms.

The paper assumes seven hours from nominal analysis initialization to operational availability, including input and dissemination. A buyer should measure actual publication lag, late or revised cycles and end-to-end decision latency. A more recent forecast is not automatically better calibrated for the event, geography or threshold that drives the application.

Marginal skill can hide implausible joint samples

The model trains on marginal continuous ranked probability score. Google explicitly says this can let the model “cheat” on covariance structure. Individual precipitation and station-head members show hexagonal patterns inherited from the mesh; PARDIG artifacts remain in the ensemble mean and are less visible in the median.

The station head also develops global warm-or-cold member bias that changes at each six-hour noise draw. Quantiles are largely cleaner, but a path-dependent simulator, grid optimizer or catastrophe workflow may consume individual trajectories rather than only medians. Those applications need spatial and temporal coherence tests, not only pointwise CRPS.

Cyclone improvement comes with under-spread

The homogeneous 2024 cyclone study reports small, consistent track and intensity improvements and larger extent gains at one-to-three-day leads. It also finds WeatherNext 3 generally more under-spread than WeatherNext 2 for intensity and extent, possibly because the larger model overfits cyclone variables.

Under-spread is a decision risk: an ensemble can improve mean error while assigning too little probability to bad outcomes. Evaluate rank histograms, coverage and threshold reliability for the exact basin and action, and keep official meteorological guidance authoritative.

Independent evaluation corroborates operation, with a caveat

Brightband’s Operational WeatherBench independently lists WeatherNext 3 in a live WeatherBench-X pipeline and exposes deterministic and probabilistic comparisons. At retrieval, the dashboard showed an August 26, 2026 12Z latest ranked cycle and regional links led by WeatherNext 3, while WeatherNext 2 held the displayed three-cycle active streak for the selected temperature metric.

Brightband’s methodology says WeatherNext 3 has no analysis of its own, so it is verified against ECMWF IFS analysis, while other models may use their own lead-zero analysis. This is transparent and useful, but the target is not identical across every row. Live rank is corroborating operational evidence, not universal proof.

The paper is candid about data and geography

Training combines ERA5 from 1959, HRES-era data, an 11-channel satellite mosaic, two precipitation products, station networks and cyclone records. The paper reports major station gaps and initial biases in places including the Andes, Himalayas and high-latitude oceans. It adds 2,000 pseudo-stations per hour filled from analysis data to reduce those biases.

That correction can improve global behavior while partially reintroducing the analysis dependency the station head was meant to escape. Validate local elevation, coast, mountain and sparse-observation slices independently. Global resolution is not evidence of local calibration.

Open source does not mean this production checkpoint is open

Google’s overview points to open-source research models, pretrained weights and notebooks. The WeatherNext 3 launch and paper do not provide a downloadable production-checkpoint hash, training code bundle or complete serving artifact for the operational system reviewed here.

Do not write “WeatherNext 3 is open source” into architecture or procurement records without binding the exact repository, revision, weights, license and parity evidence. A related research implementation and a provider-operated production forecast are different supply-chain objects.

A forecast feed needs an operational contract

Record issue time, nominal initialization, availability time, revision, variable, vertical level, grid, unit, accumulation window, ensemble member semantics and missing-value policy. Reject stale data before it enters downstream optimization. Store the provider cycle identifier with every decision so the result can be reconstructed.

Check the feed at the expected cadence and alert on lateness, schema drift, partial members and impossible values. Never silently substitute yesterday’s cycle or a deterministic median where a probabilistic threshold was expected. Make the fallback explicit: incumbent forecast, no-action state or human meteorological review.

Run a shadow evaluation around decisions

Backtest both forecasts on the historical decisions the system actually makes: renewable dispatch, routing, inventory, staffing, insurance triage or flood-model input. Freeze decision timestamps so no model receives observations unavailable at the time. Score calibration, proper scoring rules and economic loss by region, season, horizon and threshold.

Then shadow live cycles for enough time to include ordinary misses, delayed data and rare conditions. Do not optimize a threshold on the same window used to approve it. Pre-register the incumbent, acceptance margin, uncertainty method, severe-error ceiling and rollback trigger.

Test artifacts against the downstream consumer

Feed individual ensemble members—not only summary statistics—through the downstream pipeline. Probe whether mesh hexagons create false localized peaks, whether six-hour jumps trigger control actions, and whether globally biased members distort portfolio risk. Compare mean, median and exceedance products without selecting the best after seeing outcomes.

If a use case depends on coherent trajectories, require domain constraints and independent observation checks. If it only needs marginal exceedance probabilities, verify reliability at those thresholds and still monitor distribution shift. The paper’s artifact disclosure is a test plan, not a reason to dismiss the model wholesale.

Keep official warnings outside the model authority boundary

The paper labels WeatherNext an automated experimental AI system, says forecasts are provided as-is, and says they are not official severe-weather warnings. Any life-safety interface must preserve official local and national alerts as the authoritative channel.

An AI forecast may support internal analysis without becoming the sole trigger for evacuation, shutdown or public warning. Document who can override it, which source wins on conflict and how users see provenance and uncertainty.

Set rollback and re-evaluation conditions now

Rollback or fail closed when issue latency breaches the freshness budget, required ensemble members are absent, schema or units change, local calibration falls below the incumbent, rare-event misses exceed the severe-error ceiling, downstream artifact tests fail, or the access surface changes model identity.

Re-evaluate after a provider checkpoint, upstream ECMWF cycle, precipitation-head, schema or dissemination change. The paper notes that the production model includes a new ECMWF cycle; upstream distribution changes are part of the effective model version even when the WeatherNext name stays fixed.

Approve one surface and one decision

Consume WeatherNext 3 data where a named decision benefits under frozen, time-correct evaluation. Pilot before replacement. Wait for explicit WeatherNext 3 managed-inference documentation when custom generation is the requirement. Reject life-safety authority and unmonitored stale-data fallback.

The durable finding is not simply that WeatherNext 3 is more accurate. It is that a major AI model can be operational while its build surface, managed-inference surface and evidence surface remain distinct. Deployment starts by naming which one you are actually buying.

Disclosures

AccessAllGPT did not receive, use, benchmark or receive advance access to WeatherNext 3, WeatherNext 2 managed inference, Google Cloud forecast data or Brightband’s underlying database. Google, Google DeepMind, Google Research and Brightband did not sponsor, review or provide private data for this work and receive no endorsement. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Google, Google DeepMind, OpenAI, Brightband or the other organizations cited. Publication-wide relationships are listed on the disclosures page.