Which forecast is your app using? What 15 wind, surf and weather apps disclose

Why this matters now

AI weather models moved from research into service in the past two years. ECMWF's AI model (AIFS) runs every day, Google's WeatherNext models feed Search and phone weather, and the US national weather agency launched its own AI forecast in December 2025, using about 0.3% of the computing power of its standard model for a 16-day forecast. Several of the apps below now offer or blend these models.

A good global score is not the same as a good forecast at your beach. The published evaluations mostly compare AI models with ERA5, a reconstruction of past weather, rather than with real stations; a 2026 study of one AI model calls checking it against observations an important outstanding task, and its 28 km grid cannot resolve features smaller than about 100 km — the scale of a sea breeze. So the useful questions for a rider are simple: which model is this, and how often is it right here?

What we checked

We asked the same five questions of each app: does it name the models behind its forecasts; does it say it uses AI, and how; can you pick the model; does it publish accuracy results; and does it state the resolution or how often the forecast updates. We counted only what each company says on its own website, help centre, documentation or official blog; press is used once, and labelled. “Not found” means we did not find it on the official pages we could reach on 3 November 2026, not that it does not exist. AYON Ecosystem's own apps are scored on the same questions.

The results

Each cell summarises the company's own wording; the sources are listed below.

• Windy

  Names its models

  Yes — ECMWF, GFS, ICON, AROME and more

  Says it uses AI

  Yes — offers ECMWF's AI model (AIFS) for 15-day Premium forecasts

  You can pick the model

  Yes

  Publishes accuracy

  Not found (an in-app forecast-vs-observation view only)

  Resolution or update rate

  Yes

• Windguru

  Names its models

  Not verified — help page unavailable; press names GFS, ICON, WRF, AROME

  Says it uses AI

  Not found

  You can pick the model

  Not verified

  Publishes accuracy

  Not found

  Resolution or update rate

  Not verified

• Windfinder

  Names its models

  Yes — GFS; WRF/RRFS for its Superforecast

  Says it uses AI

  Not found

  You can pick the model

  Partly — Forecast or Superforecast

  Publishes accuracy

  Not found

  Resolution or update rate

  Yes, though two of its pages give different figures

• Surfline

  Names its models

  Yes — its own LOTUS wave model

  Says it uses AI

  Yes — “informed by machine learning”

  You can pick the model

  Not found

  Publishes accuracy

  Headline claim only — errors cut “by over 25%”

  Resolution or update rate

  Not found

• PredictWind

  Names its models

  Yes — ECMWF, AIFS, GFS, ICON, UKMO and its own models

  Says it uses AI

  Yes — names the AI models in its blend (AIFS, GraphCast, FengWu)

  You can pick the model

  Yes

  Publishes accuracy

  Self-published evaluation, in charts

  Resolution or update rate

  Yes

• Windy.app

  Names its models

  Yes — ECMWF, GFS, ICON, AROME, its own WRF8 and more

  Says it uses AI

  Not found

  You can pick the model

  Yes

  Publishes accuracy

  Per spot only — a model rating against a nearby station

  Resolution or update rate

  Partly

• WeatherFlow / iKitesurf

  Names its models

  Yes — ECMWF, US national models and its own WF-WRF

  Says it uses AI

  Yes — AI post-processing (“WF AI”)

  You can pick the model

  Yes, on iKitesurf

  Publishes accuracy

  Not found

  Resolution or update rate

  Not found

• Wisuki

  Names its models

  Yes — GFS 27 km and GFS Wave 50 km, on forecast pages

  Says it uses AI

  Not found

  You can pick the model

  Not found

  Publishes accuracy

  Not found

  Resolution or update rate

  Resolution only

• Surf-Forecast.com

  Names its models

  Yes — GEFS and GFS-Wave; one new model not named

  Says it uses AI

  Not found

  You can pick the model

  Yes, on Premium

  Publishes accuracy

  Not found

  Resolution or update rate

  Partly

• Stormglass

  Names its models

  Yes — ECMWF, AIFS, the US AI model (AIGFS) and more

  Says it uses AI

  Yes — offers AI models (AIFS, AIGFS) among its sources

  You can pick the model

  Yes, by API parameter

  Publishes accuracy

  Not found

  Resolution or update rate

  Partly

• Apple Weather

  Names its models

  No — lists agencies, not models

  Says it uses AI

  Not found

  You can pick the model

  Not found

  Publishes accuracy

  Not found

  Resolution or update rate

  Not found

• Google Weather

  Names its models

  Yes — WeatherNext 2 and WeatherNext 3

  Says it uses AI

  Yes — WeatherNext is an AI model

  You can pick the model

  Not found

  Publishes accuracy

  Model-level claims, plus outside evaluations

  Resolution or update rate

  Yes — 5 to 25 km, hourly

• AccuWeather

  Names its models

  No — “several dozen computer models”

  Says it uses AI

  Yes — “proprietary machine learning”

  You can pick the model

  Not found

  Publishes accuracy

  Self-run study: one city, one month

  Resolution or update rate

  Not found

• The Weather Channel

  Names its models

  Yes — its own GRAF model, blended with 100+ others

  Says it uses AI

  Yes — an “AI forecast engine”

  You can pick the model

  Not found

  Publishes accuracy

  Commissioned study, 2021–2024

  Resolution or update rate

  Relative claims only

• Ventusky

  Names its models

  Yes — ECMWF, GFS, ICON, HRRR and more

  Says it uses AI

  Not found

  You can pick the model

  Yes

  Publishes accuracy

  Not found

  Resolution or update rate

  Yes

• AYON Wind / AYON Wave

  Names its models

  Partly — names its source (Open-Meteo's best-match blend), not the models inside it

  Says it uses AI

  No

  You can pick the model

  No

  Publishes accuracy

  Graded swell calls in this library; no figures for the apps' wind numbers

  Resolution or update rate

  Update rate only

What the pattern shows

Naming models is now normal for wind and surf apps, and most let you compare them. General weather apps go the other way: Apple and AccuWeather name no models, and the big consumer apps show one blended number. AI is spreading faster than the explanation of it: 8 of the 15 say they use AI, but 4 of those — Surfline, WeatherFlow, AccuWeather and The Weather Channel — do not say which AI model or what it changes.

Accuracy is the real gap. Only 3 apps publish accuracy figures: Google, AccuWeather and The Weather Company. AccuWeather's is its own study of next-day temperatures in one city for one month (an average error of 1.50°F against 1.85°F for the national weather service), and The Weather Company's 2021–2024 study, covering 84 metrics, was commissioned by the company. PredictWind publishes a self-run evaluation in charts, Surfline a headline percentage, and Windy.app a per-spot rating. We found no wind or surf app that publishes independent verification.

• 12 of 15 name their models; 2 do not; 1 could not be verified.
• 8 of 15 say they use AI; 4 of those do not say which AI or how.
• 7 of 15 let you pick the model outright; 1 more partly.
• 3 of 15 publish accuracy figures — one self-run, one commissioned, one at the level of the model.
• 0 of the wind and surf apps publish independent verification.

Where AYON Ecosystem stands

On the same yardstick, AYON Ecosystem is in the middle, not at the top. Its methodology page names the source of its wind numbers — Open-Meteo's best-match blend of public models, read at each spot's coordinates — but not which models are inside that blend at a given spot. It states no AI forecasting, lets no one pick a model and gives no resolution. It does publish locked swell calls in this library and grades them afterwards, misses included, but it publishes no accuracy figures for the apps' wind numbers yet.

That gap is what our next study is for: testing AI and physics models against real wind stations at kite beaches, and publishing the results whether they are good or bad.

What would make this audit wrong

Companies change their pages often, and some of what we looked for may sit inside the apps rather than on public pages. Some pages could not be reached during this check, including Windguru's model help page and parts of Surfline's site, so those rows say “not verified” or “not found” rather than “no”. If your company's row is wrong or out of date, write to contact@withayon.com with the page that shows it, and we will correct the table and date the change.

References & data

1. 1.Windy — What source of weather data does Windy use? (community forum, updated Aug 2026) — Official community post listing the models on Windy (ECMWF, GFS, ICON, AROME and others) and their resolutions.
2. 2.Windy — Premium 15-day forecast for the ECMWF model using AIFS (community forum, Jul 2024) — Announcement of a Premium 15-day ECMWF forecast using AIFS, described as “advanced machine learning models”.
3. 3.Bateaux.com — Décrypter Windguru: modèles, prévisions et astuces (press, Jan 2025) — Press article listing the models Windguru shows (GFS, ICON, WRF, AROME); Windguru's own model help page could not be reached.
4. 4.Windfinder — About — Models behind Forecast (GFS) and Superforecast (WRF/RRFS), with resolutions and update times.
5. 5.Surfline — What is LOTUS? — LOTUS, Surfline's own wave model, “informed by machine learning”; errors reduced “by over 25%” against its previous model.
6. 6.PredictWind — Weather models — Model list with resolutions and update times; PWAi blend of AI models (AIFS, GraphCast, FengWu); self-published evaluation study.
7. 7.Windy.app — Weather forecast models (Jul 2024) — Model list (ECMWF, GFS, ICON, AROME, own WRF8 and others); compare mode and per-spot model rating.
8. 8.Tempest (WeatherFlow) — Tempest forecasting process (Sep 2024) — Forecast process: ECMWF and US national models as inputs, own WF-WRF, machine-learning post-processing (“WF AI”).
9. 9.Wisuki — Carcavelos forecast page — Forecast page stating GFS 27 km and GFS Wave 50 km.
10. 10.Surf-Forecast.com — Premium launches 16-day forecast and new improved modelling (Mar 2026) — GEFS ensemble and GFS-Wave in a 16-day Premium forecast; a new high-resolution model not named.
11. 11.Stormglass — FAQ — Sources include ECMWF, NOAA, Met Office, AIFS and AIGFS; automatic source selection “depending on performance”.
12. 12.Apple — WeatherKit data source attribution — Lists agencies under weather models, with no model names.
13. 13.Google — Introducing WeatherNext 3 (Sep 2026) — WeatherNext 3 in Search, Gemini and Maps; 5, 10 and 25 km, updated hourly; model-level accuracy claims.
14. 14.AccuWeather — June 2025 forecast accuracy case study — Own study: next-day temperatures in Washington, DC, June 2025; average error 1.50°F against 1.85°F for the national weather service.
15. 15.The Weather Company — TWCo widens lead as world's most accurate forecaster (Jun 2025) — ForecastWatch study of 2021–2024, 84 metrics, commissioned by The Weather Company; WxMix described as an “AI forecast engine” blending 100+ models.
16. 16.Ventusky — Help — Model list with resolutions and update intervals; model picker.
17. 17.Dirt & Data — AI weather models are here (corrected Oct 2026) — Status of AI weather models in service, including the US AI forecast launched in December 2025 at about 0.3% of the standard model's computing for a 16-day forecast; local accuracy largely untested.
18. 18.Huang et al. — Predictability of selected weather extremes with Aurora (Atmosphere, 2026) — Evaluation of one AI model against ERA5, not observations; checking against observations called an important outstanding task; a 28 km grid cannot resolve features below about 100 km.
19. 19.AYON Ecosystem — Methodology — AYON Ecosystem's own disclosure: Open-Meteo's best-match blend read at spot coordinates; forecasts update as models publish new runs, typically every few hours.

More from the research library