WeatherNext 3 Explained: Benchmarks, Hourly Forecasts, Access and Important Limits

Google DeepMind and Google Research released WeatherNext 3 on September 3, 2026 as a new operational global weather-forecasting model. It is not a general-purpose language model: there is no token context window, SWE-bench score or chat-style API to compare directly with GPT-6 Astra, Claude Fable 5.1 or other frontier LLMs. Its relevant questions are forecast skill, update latency, spatial resolution, ensemble calibration, access surface and reliability under real-world meteorological conditions.

What WeatherNext 3 actually is

The official WeatherNext 3 paper describes a probabilistic Functional Generative Network (FGN) system using an encode-process-decode architecture. It produces 64-member ensemble forecasts out to 15 days for the main 00/06/12/18 UTC cycles. Interim hourly initializations have a shorter 48-hour horizon in the operational datasets.

The model combines traditional atmospheric analysis with newer observation modalities. It consumes two analysis frames six hours apart plus recent geostationary satellite mosaics. The satellite stream is available hourly and with substantially lower latency than the underlying numerical-weather analysis, which allows WeatherNext 3 to issue a new forecast every hour rather than only every six hours.

WeatherNext 3 also changes resolution by output type rather than pretending that every variable is available at the same detail. The current system provides approximately 0.05 degree (~5 km) station-calibrated output for 2 m temperature and dew point, 0.1 degree (~10 km) gridded surface variables, and 0.25 degree (~25 km) upper-air fields. That distinction matters: the often-repeated 'five times sharper' claim applies to selected surface outputs versus WeatherNext 2's 0.25-degree grid, not to every atmospheric variable.

The paper says the model scales the latent width from 768 to 1024 and the mesh-transformer depth from 24 to 32 layers relative to WeatherNext 2. It also uses a multi-stage curriculum that progressively increases resolution from 1 degree to 0.25 degree to 0.1 degree, followed by specialized station-head training. Google does not publish a simple consumer-facing parameter count in the release material, so this article does not infer one.

Training data and sample coverage

WeatherNext 3 is trained on a mixture of ERA5, ECMWF HRES-era analysis, an 11-channel geostationary satellite mosaic, two precipitation products, surface-station observations and cyclone records. The paper extends ERA5 training back to 1959 and says the production model used for real-time evaluation was trained through June 30, 2026.

For direct surface observations, the paper combines METAR, Mesonet and ICOADS. It describes roughly 5,000 METAR airport stations and around 15,000 Mesonet stations in 2024, with the number changing over time. Across the observation sources, it reports about 3,000 observations per hour on average in 2024. A temporally consistent 5% of METAR and Mesonet stations is held out from training for evaluating spatial generalization.

That dataset is not geographically uniform. The authors explicitly report major observation gaps and initial model biases in sparsely observed regions including the Andes, Himalayas and some high-latitude ocean areas. They mitigate this partly with pseudo-station samples, but the gap remains an important limitation when interpreting global claims.

How the benchmark evidence was produced

The strongest release evidence comes from the paper rather than a single headline score. Google evaluates WeatherNext 3 using Continuous Ranked Probability Score (CRPS) as its main probabilistic metric and also uses Brier scores, reliability diagrams, pooled CRPS and task-specific cyclone measures. Lower CRPS is better.

The main non-real-time evaluations cover the full year of 2024 using a model trained only through the end of 2023, which helps reduce train-test leakage. The production checkpoint is evaluated separately on a six-week quasi-real-time period from July 1 to August 11, 2026 against operational WeatherNext 2 and ECMWF AIFS ENS feeds. The authors themselves caution that six weeks is a relatively small sample and that small differences should not be over-interpreted.

For precipitation, Google's September 3 launch says WeatherNext 3 shows CRPS improvements of up to 60% against IMERG, about 30% against MRMS, and about 10% against rain-gauge observations for early lead times, depending on the evaluation. These are not one universal '60% more accurate' number. They use different observation targets and lead-time conditions. The developer benchmark page gives a more conservative summary of up to roughly 50% improvement in Brier score and CRPS versus numerical-weather baselines in IMERG-based comparisons.

The paper also limits its precipitation evaluation to moderate accumulations of up to 4 mm per six hours because heavier events become too sparse for reliable statistics in that setup. Therefore the published precipitation benchmark should not be interpreted as proof of equal gains for rare extreme rainfall.

The hourly-initialization experiment is especially useful because it isolates a practical operational advantage. Google compares the hourly system with the same model constrained to six-hourly cycles and estimates a 2-3 hour effective lead-time advantage for precipitation at short ranges. For its latency-adjusted analysis, the paper assumes roughly seven hours from nominal analysis time to operational availability, including data latency, model execution and dissemination.

What independent evaluation says

Google points to Brightband's Operational WeatherBench as independent live evidence. Brightband runs a WeatherBench-X-based operational leaderboard covering AI and numerical models. Its methodology currently includes WeatherNext 3 as a machine-learning ensemble initialized from ECMWF IFS analysis.

This is useful independent corroboration, but the Brightband methodology contains a major comparability caveat. By default, operational systems are verified against their own analysis. WeatherNext 3 does not publish an analysis of its own, so Brightband substitutes ECMWF IFS analysis as its verification target. Brightband also currently omits WeatherNext 3 precipitation from cross-model scores because the model's precipitation output is trained against a satellite-based target and is not straightforwardly comparable with the NWP/reanalysis precipitation fields used by other systems.

That means a headline such as 'WeatherNext 3 is number one' should be read as a result inside a specific live benchmark configuration, not universal proof that the model is superior for every variable, region, severe-weather event or application. The independent live leaderboard is stronger evidence than a vendor-only claim, but it still has scope and target-definition limits.

Access, pricing and deployment are easy to misunderstand

Google's WeatherNext 3 quick start currently distributes forecast data through Google Cloud Storage, BigQuery and Earth Engine. Access is allowlisted on request, and the documentation says requests are typically reviewed within 5-7 business days. A paid Google Cloud contract is not required simply to request access.

The raw Google Cloud Storage surface exposes the full 64-member ensemble and pressure-level data, while BigQuery and Earth Engine expose precomputed surface statistics such as mean and percentiles. Main six-hour cycles cover up to 360 hours (15 days); interim hourly cycles cover 48 hours.

There is no published per-token or per-forecast WeatherNext 3 inference price comparable with an LLM API. Cloud platform usage can still incur costs. For example, the full-ensemble Cloud Storage bucket is Requester Pays and may generate network-egress charges, while BigQuery and Earth Engine have their own platform cost and quota models.

A second important access boundary is that Google's current managed inference specification still names WeatherNext 2 as the model accessible on the Gemini Enterprise Agent Platform. In other words, access to WeatherNext 3 operational forecast arrays should not be silently described as on-demand managed WeatherNext 3 model inference.

The paper reports that generating one ensemble-member forecast takes about 6.3 minutes using four TPUv5p chips. It also reports substantial training compute across TPUv4 and TPU7x. That figure is a research-system inference measurement, not a promise of end-user API latency.

Safety and practical limits

WeatherNext 3 forecasts are explicitly experimental. Google says they are not official severe-weather warnings and should not replace national meteorological agencies or emergency authorities.

The paper documents several limitations beyond sparse station coverage. It reports under-dispersion in some cyclone attributes, discusses artifacts in joint forecasts and notes that some output variables required clipping during inference to prevent unrealistic values from accumulating. The real-time comparison window is also only six weeks, so long-season and rare-event performance still needs broader independent observation.

For developers, the practical conclusion is to validate WeatherNext 3 against the exact decision you care about rather than relying on a global rank. Agriculture, renewable-energy forecasting, logistics and local temperature applications may value its hourly refresh and surface resolution differently from aviation, tropical-cyclone risk or extreme-precipitation workflows.

Why SWE-bench and LLM leaderboards do not belong here

WeatherNext 3 has no meaningful SWE-bench Verified, SWE-bench Pro, Terminal-Bench, token context-window or chat-tool-use score. Those benchmarks measure software engineering or language-agent behavior, while WeatherNext 3 predicts physical weather fields. Mixing them would create a false ranking across incomparable systems.

The relevant comparison set is WeatherBench-style meteorological evaluation, ECMWF operational baselines, independent live forecast verification and application-specific backtesting.

Public feedback

This review checked for directly attributable public discussion in addition to Google and Brightband documentation. No sufficiently reliable first-hand X post was found that would justify a broad user-sentiment claim. That absence is reported rather than replaced with an invented consensus. Early reporting is largely focused on the launch and benchmark results, so operational user evidence is still thin.

Bottom line

WeatherNext 3 is a meaningful AI-model release because it combines global probabilistic forecasting with hourly satellite-fed updates, higher-resolution surface outputs and a 64-member ensemble. The evidence is stronger than a typical launch claim: Google provides a detailed 2024 evaluation, a six-week 2026 operational comparison and Brightband supplies independent live benchmarking.

The limitations are equally important. The largest precipitation gains depend on the target and lead time, extreme rainfall is not fully covered by the reported benchmark, observation density varies globally, Brightband has verification-target caveats, and the public Cloud offering is forecast-data access rather than a general WeatherNext 3 managed-inference endpoint. For consequential weather decisions, official meteorological warnings remain authoritative.