DeepMind Opens WeatherNext — Cost Impact
DeepMind open-sources WeatherNext after gaining a day of cyclone forecast accuracy. See access options, compute needs, and deployment advice.
By AI Pricing Guru Editorial Team
AI Pricing Guru articles are maintained by the editorial workflow behind the site: daily pricing snapshots, provider source checks, and review passes for model launches, subscription limits, and billing changes.
TL;DR
- DeepMind says WeatherNext Cyclones gains more than 24 hours of lead-time accuracy across cyclone track, intensity, and wind-structure forecasts.
- The code and model weights are now open, including a Mini version that runs on a single TPU or GPU and a free Colab path.
- Open weights remove a model-access fee, but teams still pay for accelerators, weather inputs, storage, operations, and expert validation.
- WeatherNext supports professional forecasters; it does not replace official warnings from meteorological agencies.
Google DeepMind has open-sourced WeatherNext 2 and WeatherNext Cyclones after reporting a step change in tropical-cyclone forecasting. In a peer-reviewed Nature paper, the team says its three-day forecasts are as accurate as previous systems’ two-day forecasts on average.
That extra day covers track, intensity, and wind structure—not only where a storm may go. The release changes the cost question from “Can we license the model?” to “Can we operate and validate it reliably?”
WeatherNext is not a Gemini chat model, so it adds zero rows to the canonical pricing dataset and pricing API. A zero-dollar API row would incorrectly turn open-source access into a hosted service claim. Our maintained Google AI pricing page now separates WeatherNext access and infrastructure costs from Gemini token pricing.
What DeepMind announced
WeatherNext Cyclones was evaluated on historical storms from 2023 and 2024. DeepMind reports more than 24 hours of average lead-time advantage over leading weather models, which it describes as roughly a decade of meteorological progress.
The model combines global atmospheric dynamics with expert-curated cyclone observations. Training used nearly 20 terabytes of weather data and the IBTrACS archive of almost 5,000 storms. It predicts a cyclone’s path, strength, and wind field in one system, bridging work traditionally split between coarse global and high-resolution local models.
DeepMind says one 15-day forecast takes less than a minute on a TPU. Its operational workflow now generates 1,000 possible scenarios per cyclone, up from 50, to better represent rare outcomes such as rapid intensification.
Access and cost comparison
DeepMind did not announce a universal per-forecast API rate. Instead, it released research code, pretrained weights, forecast feeds, and several ways to run or consume the system.
| Deployment path | Access status | Compute requirement | Best fit |
|---|---|---|---|
| WeatherNext Cyclones Mini | Open code and weights; free Colab runtime available | Single TPU or GPU; DeepMind says a P100-class GPU should handle Mini inference | Evaluation, teaching, and local prototypes |
| Full WeatherNext 2 or Cyclones | Open code and weights | TPU recommended; an H100-class GPU is required for sufficient memory on non-Mini models | Research and operational pilots |
| Published forecast feeds | Available through Weather Lab, Google Cloud, and Open-Meteo | Provider-managed | Applications that need outputs without model operations |
The repository licenses its code and notebooks under Apache 2.0 and other released materials under CC BY 4.0. Input datasets can carry separate terms, so open model access does not automatically make an end-to-end service unrestricted or costless.
Google also publishes daily WeatherNext forecast datasets through BigQuery, Earth Engine, and Cloud Storage. Those access routes do not create a WeatherNext model price: normal query, storage, processing, and transfer charges for the chosen Google Cloud surface can still apply. Open-Meteo’s WeatherNext endpoint is a separate third-party service with its own usage terms.
Pricing impact: who benefits and who pays
Meteorological agencies, universities, nonprofits, insurers, energy planners, and disaster-response teams gain a credible starting point without negotiating access to proprietary weights. The Mini checkpoint also lowers the hardware barrier for validation.
The remaining bill moves to infrastructure and operations: accelerator time, initial-condition data, forecast storage, ensemble generation, monitoring, retraining, and qualified meteorological review. A production team generating 1,000 scenarios will have a very different cost profile from a researcher testing one Mini forecast in Colab.
Google Cloud and Open-Meteo feeds may be simpler than self-hosting, but their service, data-transfer, and storage terms must be priced separately. WeatherNext is not a Gemini API model, so it should not be budgeted by tokens or added to a language-model invoice.
Labs availability blocker
WeatherNext is publicly available, but it does not fit the current Cost-per-Task Labs leaderboard. The 49-task harness measures deterministic language-model answers, accepted token usage, and official inference rates. WeatherNext takes gridded atmospheric initial conditions and produces weather fields and cyclone trajectories; text prompts would not test the released model.
A valid WeatherNext benchmark needs a pinned checkpoint and serving stack, fixed licensed initial conditions, a cyclone tracker, held-out verification data, and predeclared track, intensity, wind, and probabilistic skill metrics. It also needs accelerator-hours, ensemble size, wall-clock time, storage, and transfer logs so cost per forecast can be calculated at a fixed quality threshold.
That weather-specific harness is the explicit blocker. Repeating Google’s paper result as a Labs score would imply independent reproduction that has not happened.
What teams should do now
- Reproduce a small historical storm with WeatherNext Cyclones Mini before provisioning full-model hardware.
- Score track, intensity, wind extent, latency, and total infrastructure cost against the forecasting system already in use.
- Test ensemble sizes that match the decision being supported; do not assume the 1,000-member research workflow is necessary for every product.
- Pin the repository release because DeepMind labels this research code and does not guarantee API stability.
- Keep official weather-service alerts as the public safety authority and position AI output as decision support.
For adjacent AI workloads, compare current Google model pricing with OpenAI pricing and use the token calculator. Those tools cover language-model costs, not WeatherNext infrastructure. Our API versus self-hosting break-even guide provides a framework for counting hardware, utilization, and engineering overhead.
What this means
WeatherNext’s most important commercial effect is not a cheaper token rate. It is that a state-of-the-art forecasting system can now be inspected, tested, and adapted without buying model access first.
That broadens participation, but it does not remove deployment risk. Teams should judge WeatherNext by forecast skill and cost per useful decision, validate it with professional forecasters, and never substitute its output for an official warning.
Sources: Google DeepMind’s official WeatherNext announcement, the peer-reviewed Nature paper, the open WeatherNext repository, Google’s official model documentation and forecast-access guide, and the 2025 NHC verification report. Access, licensing, and pricing status checked August 8, 2026.