- DeepMind reports GenCast beating the European ensemble on 97.2 percent of 1,320 variable and lead time combinations, rising to 99.8 percent beyond 36 hours.
- The speed difference is the part that changes decisions: about 8 minutes for a 15 day ensemble on one cloud chip, against hours on a supercomputer.
- These models did not replace the forecasting system. They replaced one step inside it, and still start from a state built by the physics based assimilation of around 60 million observations.
- Europe's forecasting centre took its own AI model operational on 25 February 2025 and reports gains of up to 20 percent on some measures, including cyclone tracks.
- For a business that plans around weather, the useful change is cheap ensembles, which means probabilities rather than a single number to plan against.
97.2 percent. That is the share of 1,320 tested combinations of variable and lead time on which DeepMind's GenCast model beat the ensemble forecast of the European Centre for Medium Range Weather Forecasts, the system widely treated as the best operational forecast in the world. Beyond 36 hours ahead, the figure goes to 99.8 percent.
Those numbers are real and they are published. They are also routinely misread, in a way that matters to anybody who plans a business around the forecast. The headline version says AI beat the weather services, and machine learning weather prediction is now treated as a settled win. What actually happened is narrower, more interesting, and much more useful to know.
What exactly did the AI models beat?
One step of the forecasting pipeline, measured against the same step done the traditional way. The DeepMind write up of GenCast, published on 4 December 2024, reports the 97.2 percent figure across 1,320 variable and lead time combinations against ENS, with the win rate rising to 99.8 percent at lead times beyond 36 hours. The earlier deterministic model, GraphCast, announced on 14 November 2023, was more accurate than the HRES system on more than 90 percent of 1,380 test variables and lead times, and on 99.7 percent of test variables in the troposphere.
Worth noting if you go to the primary literature: the arXiv version of the GenCast paper states greater skill than ENS on 97.4 percent of the 1,320 targets evaluated, while the blog post states 97.2 percent. The two figures differ slightly and both come from the same team. Either way the claim is the same shape, and anyone quoting a single decimal should say which document they took it from.
The speed figures are less disputed and arguably more consequential. GenCast produces a 15 day global ensemble forecast at 12 hour steps and quarter degree resolution, covering more than 80 variables, in about 8 minutes on a single cloud chip, with ensemble members generated in parallel. GraphCast produces a 10 day forecast in under a minute. The physics based equivalents take hours on a supercomputer with, in the case of ENS, tens of thousands of processors.
Why can these models not replace the weather service?
Because they do not touch the hardest part of the job, which is working out what the atmosphere is doing right now. A forecast model needs a starting state: a complete, physically consistent picture of the global atmosphere at one moment. Producing that is called data assimilation, and it means fusing millions of irregular observations from satellites, balloons, aircraft and surface stations into a single coherent field.
That step is still done by the physics based system, and the AI models are initialised from its output. Europe's own announcement is explicit. When ECMWF took its Artificial Intelligence Forecasting System operational on 25 February 2025, it stated that the AIFS uses the same initial conditions for its forecasts as the conventional IFS, and that those conditions combine a previous short term forecast with around 60 million quality controlled observations. The AI model runs with a grid spacing of 28 km and, by ECMWF's own account, outperforms state of the art physics based models on many measures, with gains of up to 20 percent including on tropical cyclone tracks.
So the honest sentence is this: machine learning has taken over the forecast step and not the observation and assimilation pipeline, and it could not run without it. There is also a training dependency. Both DeepMind models learned from ERA5, the reanalysis dataset ECMWF built by running its physics model over four decades of historical observations. The AI models are, in a real sense, distillations of the physics system's own record of the atmosphere.
None of that diminishes the result. It does change what you should expect if the public forecasting infrastructure were ever cut: the AI models would degrade with it rather than carry on regardless.
Why is the advantage bigger further ahead?
Look again at the two GenCast numbers. The win rate against the European ensemble is 97.2 percent across all tested targets and 99.8 percent at lead times beyond 36 hours. The advantage is not uniform. It grows with distance.
The reason is worth understanding, because it tells you where the AI contribution actually sits. Very short range forecasting is dominated by the quality of the starting state. If you know precisely what the atmosphere is doing now, the next few hours follow fairly directly, and the model has less room to add or lose skill. That starting state is produced by the assimilation system, which both approaches share. So at twelve hours the two are working from the same advantage and the difference is small.
Further out, the starting state matters less and the propagation of errors matters more, which is the part the model does. That is where a method trained on four decades of how the atmosphere actually evolved appears to have an edge over integrating equations forward step by step. For a business this maps onto something practical: the improvement you are benefiting from is mostly in the two to ten day window, which happens to be exactly the window in which ordering, staffing and scheduling decisions get made.
How do the two approaches compare on the things that matter?
This table is assembled from the figures above rather than from any single source, because no one source sets the two side by side.
| Dimension | Numerical weather prediction | Machine learning forecast models | What follows for a user |
|---|---|---|---|
| Measured skill, medium range | The reference standard, and still the benchmark everyone reports against | Better on 97.2 percent of 1,320 tested targets for GenCast, over 90 percent of 1,380 for GraphCast | The AI forecast is not a cheap approximation; it wins on the published scores |
| Time to produce one forecast | Hours on a supercomputer with very many processors | About 8 minutes for a 15 day ensemble, under a minute for a 10 day deterministic run | Forecasts can be refreshed and re run far more often |
| Ensemble size | Expensive, so the number of members is rationed | Cheap and parallel, so large ensembles become practical | Probabilities get better, which is what a planner actually needs |
| Dependence on observations | Performs the assimilation of around 60 million observations itself | Takes the starting state from that same assimilation | The AI model is downstream of the public infrastructure, not independent of it |
| Physical consistency | Conserves quantities by construction, since it solves the equations | Learned from data, so consistency is not guaranteed by design | Treat unusual output with more care, especially far outside past patterns |
| Extreme events | Long operational track record and known biases | DeepMind reports better prediction of extreme weather and cyclones; behaviour far outside training remains an open question | Good news on the common extremes, unproven on the record breaking ones |
What is reanalysis, and why does it matter here?
Reanalysis is the historical record of the atmosphere, reconstructed after the fact. Take every observation available for a past date, feed it through a modern forecasting model, and produce a complete, consistent global picture of that day, far better than anything available at the time. ERA5, built by ECMWF, does this across roughly four decades.
Both DeepMind models were trained on it. That matters for two reasons a business reader can use. First, it explains why these models are good: they have seen how tens of thousands of real atmospheric situations actually unfolded, which is a different kind of knowledge from solving equations. Second, it bounds what they have seen. A learned model knows the last forty years of weather. Situations with no close analogue in that record are, by construction, the ones it has least basis for, and a warming atmosphere produces more such situations rather than fewer.
This is the technical root of the physical consistency caution in the table above, and the one place where the honest answer is that nobody yet knows how these models behave in conditions with no precedent in the record. It is also why the forecasting centres are running both systems rather than switching off the one that solves the equations.
What is the difference between a deterministic and an ensemble forecast?
A deterministic forecast gives you one answer: 14 degrees and dry on Thursday. An ensemble runs the model many times from slightly different starting states and slightly different physics, and gives you a spread of answers: 40 of 50 runs dry, the other 10 with rain arriving late afternoon.
The spread is the product. Weather is chaotic, which means tiny differences in the starting state grow into large differences in outcome, and the growth rate varies by situation. Some weeks the runs agree closely out to day six. Other weeks they diverge by day two. A single number hides which kind of week you are in, and that distinction is usually more valuable than the central estimate.
This is why cheap compute changes the useful output rather than just the cost. GenCast is a probabilistic model and its ensemble members run in parallel, so a larger ensemble stops being a luxury. Larger ensembles sharpen the probability of the thing you care about: not the temperature, but the probability of rain between 10am and 4pm on Saturday, which is the number a market stall actually plans against.
What should a business do differently with this?
Plan against probabilities and stop planning against the headline number. If your revenue moves with the weather, you are already making decisions under uncertainty and almost certainly pretending otherwise.
Start by writing down the decision rather than the forecast. A garden centre decides how many trays to put outside. An outdoor events business decides on the marquee by Wednesday, with a cancellation fee attached. A builder decides whether to pour on Friday. Each has a threshold: a rainfall amount, a wind speed, a temperature. Once the threshold is explicit, a probability becomes actionable, because you can compare the cost of being wrong in each direction.
Then set the rule by cost asymmetry. If an unnecessary marquee costs a few hundred and a soaked event costs the contract, you act at a 30 percent chance of rain. If the costs are symmetric you act at 50. This is ordinary expected value arithmetic and it is the whole of what a probability is for. Our piece on choosing a forecasting method you can actually check makes the same argument for demand rather than weather, and the structure of the decision is identical.
Here is the arithmetic done once, with numbers you can replace with your own. An outdoor event, marquee hire 400, revenue 6,000, and a washout without cover costs you the event. The forecast says 30 percent chance of the rain threshold being crossed. Hire the marquee and you are out 400 with certainty. Skip it and your expected loss is 0.30 multiplied by 6,000, which is 1,800. You hire the marquee, and you would still hire it at a 10 percent probability, because 600 still exceeds 400. The break even sits near 7 percent. Most owners would have called 30 percent a reason to risk it, which is exactly the error that explicit arithmetic prevents.
Run the same calculation where the asymmetry reverses and the answer flips. Trays of bedding plants put outside cost almost nothing to move back in, and leaving them out through hail destroys them, so you act on a low probability. A delivery round rescheduled for wind costs a full day of driver time, so you act on a high one. Nothing here is weather expertise. It is the cost of each mistake, written down once per decision, and then reused every week. Stock decisions work the same way, which is the ground covered in reordering stock when you have thin sales history, and the weather sensitive version simply adds a second probability to the same sum.
Two practical cautions. First, the models you can reach through a consumer app are not necessarily any of the ones discussed here, and app providers blend sources without telling you which. If a decision is worth money, get the forecast from the national service or a provider that names its model and publishes its verification. Second, a probability is not a promise about one day. A 30 percent chance of rain that does not rain was not wrong. Judge a forecast over a season, by whether the 30 percent days rained about three times in ten.
Does any of this reach the app on your phone?
Indirectly, and not in a way you can see from the interface. Consumer weather apps are resellers. They take model output from national services and commercial providers, blend it, and present one number with an icon. Which models went into that number is rarely stated, and the blend changes without announcement.
The consequence is that the improvement described here may already be in your phone, or may not, and the app will look identical either way. For decisions worth real money the fix is to go one layer up. National meteorological services publish their own forecasts, including probabilistic products, and the European centre publishes verification scores showing how its systems actually performed. A provider that names its model and shows its verification record is telling you something. A provider that shows a sun icon and a single temperature is telling you nothing about reliability.
One more habit worth adopting if weather drives your week. Record the forecast you acted on, not just the outcome. Three months of that gives you a personal verification record for the source you use, at the lead time you actually decide on, which is far more useful than any published global forecast skill score. A forecast provider that is excellent at day seven and mediocre at day two is the wrong tool for a business that commits on Wednesday for Saturday, and you cannot learn that from a benchmark table. Our weather exposed readers in farming hit this first, which is part of why the buying questions in what farm technology actually delivers in the field insist on measurement under your own conditions rather than the vendor's.
If your weather exposure runs through what you sell rather than where you work, the same probabilities belong in the shopfront: stock levels, pre orders and delivery promises that move with the forecast instead of against it. That is a build question as much as a planning one, and our AI store builder exists for the part where the plan has to become a working page.
Where is this heading, and what stays uncertain?
The direction of travel is clear enough: the forecasting centre that built the reference system now runs an AI model operationally alongside it, which settles the question of whether this is a serious method. The research frontier has moved to the part that is still physics, which is deriving forecasts directly from observations rather than from an assimilated starting state. If that works, the dependency described above weakens.
Two honest uncertainties remain, and both are acknowledged in the field rather than hidden. Behaviour on events with no close precedent in the training record is not well established, which is precisely the case a business cares most about. And physical consistency is not guaranteed by construction in a learned model the way it is in a model that solves conservation equations, so unusual output deserves a second look rather than automatic trust.
For most businesses none of that is load bearing. The practical change already happened: forecasts beyond about a day and a half got measurably better, they got much cheaper to produce, and the cheapness is arriving as probabilities. The skill is in using them, which is a planning habit rather than a technology.
The short version
AI weather forecasting is real, measured and already operational at the institution that sets the standard. It beat the incumbent on the overwhelming majority of tested targets and it runs in minutes instead of hours. It did not replace the weather service, because it still starts from a picture of the atmosphere that only the observation and assimilation system can build, and it learned from that system's own archive. The gain you can use is cheap probability. Write down your threshold, price the mistake in both directions, and let the spread tell you how confident to be.