Evidence on influence. Perspective on progress.Independent · Source-linked · AI-assisted research
Operational usePossibilitiesWeather and resilience

AI forecasts are in service. Here is what improved.

ECMWF reported gains on forecast measures and far less energy per forecast at AIFS's 2025 launch. Its May 2026 update added AI wave and snow-cover forecasts.

The short version

  • ECMWF launched operational AIFS Single forecasts on February 25, 2025, followed by the ensemble on July 1.
  • Both systems were upgraded on May 12, 2026. Their output is available under attribution-based reuse terms.
  • At the original launch, ECMWF reported gains of up to 20% on measures including cyclone tracks, and about 1,000-fold lower energy use for making a forecast.

Why it matters. The evidence now includes a working service, reported forecast-skill gains and lower computation energy—not only a research promise.

Paper-cut illustration of a storm swirling over a coastline.
AI-generated editorial illustration; not a weather forecast or a specific storm.

Better forecasts are a measurable result

At its February 2025 operational launch, ECMWF reported AIFS improvements of up to 20% on measures including tropical-cyclone tracks, alongside roughly 1,000-fold lower energy use for making a forecast.

Those comparisons have a specific scope. They are not a 20% improvement in every weather prediction or a measurement of the entire system’s lifecycle energy. The launch compared models with different grid spacings: about 28 km for AIFS and 9 km for the physics-based IFS.

More than one forecast, and more things forecast

The operational service offers both a single forecast and an ensemble of possible outcomes. In May 2026, AIFS v2 added wave and snow-cover forecasts. The same release upgraded IFS separately; not every improvement described there belongs to the AI model.

Users can obtain forecast output under the published reuse terms. That makes the capability available for weather-dependent decisions, rather than confining it to a paper. The records do not establish how many lives or dollars any particular forecast saved.

The supporting record

Operational recordClaim 1

ECMWF's February 2025 launch announcement reports forecast gains of up to 20% on measures including cyclone tracks and about a 1,000-fold reduction in energy used to make a forecast.

Operational recordClaim 2

The May 12, 2026 AIFS v2 release introduced ECMWF's first data-driven wave and snow-cover forecasts.

Operational recordClaim 3

ECMWF dates operational AIFS Single to February 2025 and AIFS Ensemble to July 2025.

Operational recordClaim 4

ECMWF documents May 2026 version-2 upgrades and CC BY 4.0 terms for AIFS output.

What we do not yet know

  • The downstream benefit attributable to a particular forecast in a particular decision.
  • Performance under every region, variable, and extreme-event condition.

Original documents, with context

Sources you can inspect

  1. 1. Operational documentation · ECMWF

    AIFS Machine Learning data

    Documents operational forecasting and access. Avoided damages and improved decisions require separate outcome studies.

    Publication: Operational history and current data access · Checked: 2026-09-07
    Location in source: Operational dates; model versions; use of output

    Source record: ecmwf-aifs
  2. 2. Operational announcement · ECMWF

    AIFS operational launch: forecast skill and energy use

    Reported gains depend on the forecast measure. The energy comparison concerns making a forecast, not training or the entire service lifecycle.

    Publication: 2025-02-25 · Checked: 2026-09-07
    Location in source: Opening results; first operational model; 28 km versus 9 km grid-spacing discussion

    Source record: ecmwf-aifs-launch-results
  3. 3. Operational update · ECMWF

    AIFS v2 adds snow and wave forecasts

    The joint release also upgrades the physics-based IFS. IFS-only improvements must not be attributed to the AI model.

    Publication: 2026-05-12 · Checked: 2026-09-07
    Location in source: AIFS v2 paragraphs and snow-cover comparison

    Source record: ecmwf-aifs-v2-results

Corrections & additional evidence · support@aipublicopinion.com