Weather Next 3 arrived on AIxploria in the early hours of 6 September 2026, carrying a “new AI tool” badge, a Free tag, a 4.4-star rating and a rank badge reading “#41 in Research & Science”. The card is short — a single 207-character sentence, with no editorial review behind it. That brevity is unusual for the directory, and it makes the few things the card does assert worth checking carefully against what Google actually shipped.
We pulled the card, its predecessor’s card, the directory’s Research & Science category listing and every public Google source behind the model: the announcement post, the developer model guide, the research and benchmarks page, the access quick start and the arXiv paper.
The picture that comes back is not the usual one. Most of what the listing says is accurate, and one figure it reports is scoped more carefully than Google’s own consumer blog post manages. The problem is the rank the directory gives it, and the single biggest fact the card leaves out. For more context on how these releases fit together, see our AI models and tools hub.
This is the eighth article in our series checking AIxploria listings against vendor documentation, and the third in a row where the rank badge turned out to be the most revealing thing on the page. Here the badge is not wrong. It is working exactly as designed — and what it correctly reports is that the directory ranks Google’s newest flagship fourteen places below the model Google has just told developers to stop using.
Table of contents
- What Weather Next 3 Actually Is
- What the Weather Next 3 Card Says and How Fast It Landed
- The Rank Badge Puts Weather Next 3 Behind Its Own Predecessor
- Weather Next 3 and the Ranked Lists: a Fourth Data Point
- Fact-Checking Every Claim on the Weather Next 3 Card
- The Biggest Omission: Weather Next 3 Already Ships in Search and Maps
- What “Free” Really Means for Weather Next 3
- Where Google’s Own Numbers Disagree With Each Other
- Weather Next 3 Against the Alternatives on Its Own Card
- What the Weather Next 3 Listing Means for Your Business
- How to Read Any AIxploria Card After Weather Next 3
- Frequently Asked Questions
- References and Further Reading
What Weather Next 3 Actually Is
Google announced the model on 3 September 2026 at 15:00 UTC, under the name WeatherNext 3 — one word, no space. The arXiv paper, “WeatherNext 3: Increasing resolution and performance of global weather models with raw observations” by Rasp and 20-plus co-authors, was submitted the same day.
The architecture behind the listing
Weather Next 3 is a Functional Generative Network mesh transformer that produces a 64-member probabilistic ensemble. Its headline change over the previous generation is what it eats. Earlier AI weather models, including WeatherNext 2, trained and initialised exclusively on numerical weather prediction analysis data, which carries a six-hour lag and passes its own biases straight into the forecast.
The new model ingests a live global geostationary satellite mosaic as a direct model input, alongside ECMWF HRES analysis. Because the satellite feed refreshes continuously, the model re-initialises every hour — 24 runs a day, against six for its predecessor. Treating raw satellite imagery as a first-class input pushes the problem closer to computer vision than to classical physics simulation, and it is the reason the hourly refresh is possible at all.
Where the 5 km headline really applies
This is the detail most coverage flattens, and the one place the AIxploria listing is unusually precise. Weather Next 3 does not produce everything at 5 km. It emits three resolutions from a single forward pass, and only the narrowest slice is the headline number.
| Resolution | Variables | Availability |
|---|---|---|
| 0.05° (~5 km) | Station-trained 2 m temperature and dew point only | All initialisations |
| 0.1° (~10 km) | Surface wind (10 m and 100 m), pressure, sea surface temperature, cloud layers, solar radiation, 1-hour precipitation | All initialisations |
| 0.25° (~25 km) | 3D atmospheric fields across 13 pressure levels | 6-hourly cycles only, Cloud Storage |
The card’s phrasing is “a resolution of up to 5 km for temperature and humidity”. Dew point is a humidity variable, and temperature plus dew point is exactly the pair the 0.05° station head produces. The listing scoped the claim correctly. Google’s own announcement blog is looser, describing the model as “roughly five times sharper” overall without noting that the 25 km tier did not move at all.
What the Weather Next 3 Card Says and How Fast It Landed
The listing went live at 02:11:30 UTC on 6 September and was edited eleven minutes later. Everything the directory publishes about the model sits in the register below.
| Field | Value on the card |
|---|---|
| Name as listed | Weather Next 3 (two words) |
| Rating | 4.4 / 5 from 7 votes |
| Upvotes / views | 51 / 4,425 |
| Rank badge | #41 in Research & Science |
| Pricing tag | Free (price 0, marked accessible for free) |
| Verification | Verified Tool (not Gold Verified) |
| Categories | Latest AI, Research & Science |
| Alternatives listed | 43 |
| Editorial review | None — one 207-character blurb |
The upstream lag, measured
Google announced on 3 September at 15:00 UTC. The card appeared on 6 September at 02:11 UTC. That is a gap of 2 days and 11 hours — the third-fastest turnaround we have measured on this series, and evidence the directory is watching Google’s newsroom closely.
| Tool | Vendor launch | Card published | Lag |
|---|---|---|---|
| Gemini 3.8 Flash | 2 Sep 2026 | 3 Sep 2026 | 1 day |
| Claude Fable 5.1 | 1 Sep 2026 | 3 Sep 2026 | 2 days |
| Weather Next 3 | 3 Sep 2026 | 6 Sep 2026 | 2.5 days |
| GPT-6 Astra | 3 Sep 2026 | 6 Sep 2026 | 3 days |
| Lyria 3.5 | 29 Jul 2026 | 6 Sep 2026 | 39 days |
On speed of cataloguing, the directory does well. The trouble starts once you ask where the newly catalogued model was actually filed.
The Rank Badge Puts Weather Next 3 Behind Its Own Predecessor
Every card in this series carries a rank badge, and following that badge’s own link has now been worth doing three times running. On the GPT-6 Astra card, the badged category simply did not contain the model. On the Lyria 3.5 card, the badged slot was occupied by the model’s predecessor. This one is different again, and it is the cleanest of the three.
What the badge claims
The Weather Next 3 card is badged “#41 in Research & Science”, linking to the directory’s Research & Science category. The WeatherNext 2 card — still live, published 18 November 2025 — is badged “#27 in Research & Science”.
What the category listing actually shows
We walked the category. It runs to seven pages of twelve cards, and the first three pages are served as plain HTML that can be read directly. Across those 36 positions, Weather Next 3 does not appear once. Position 27 is occupied by WeatherNext 2, exactly as its own badge claims.
That last detail matters more than the absence does. It means the badge is not decorative and not broken: it is a real ordinal into this list, and we verified it against a known value. So the #41 on the Weather Next 3 card should be read as a genuine claim about where the directory places the model — fourteen positions behind the model it replaces.
The honest limit of this check
Pages four through seven sit behind a Cloudflare challenge and could not be read, so we cannot confirm what occupies slot 41 or show Weather Next 3 sitting in it. What is verified is narrower and still damaging: the model is absent from the top 36, its predecessor holds 27, and the badge system demonstrably works. We are not claiming the number is fabricated. We are pointing out what it says.
Google could not be clearer about which model it wants people using. The WeatherNext 2 developer guide now opens with a banner: “WeatherNext 2 is an older model. We recommend WeatherNext 3 for all new projects.” A dedicated model deprecation guide covers migration. The directory’s ordering runs precisely opposite to the vendor’s own instruction.
Weather Next 3 and the Ranked Lists: a Fourth Data Point
Alongside category pages, AIxploria maintains curated lists — a Top 100 and an “Ultimate List” — which run on a different clock from the tool pages. This series has been measuring that gap since 3 September, and the numbers below are occurrence counts taken from the page source.
| Model | Card age | Top 100 | Ultimate List |
|---|---|---|---|
| Weather Next 3 | ~10 hours | 0 | 0 |
| GPT-6 Astra | ~8 hours | 0 | 0 |
| Lyria 3.5 | ~9 hours | 0 | 0 |
| Claude Fable 5.1 | 3 days | 16 | 10 |
| Gemini 3.8 Flash | 3 days | 16 | 10 |
| WeatherNext 2 (predecessor) | 10 months | 4 | 0 |
What the control models tell us
The two three-day-old cards read 16 and 10 for the third consecutive measurement, unchanged across the day. That stability is what makes the zeros interpretable rather than alarming: a card takes roughly three days to reach the ranked views, and a same-day zero is normal. Weather Next 3 is inside that window.
The uncomfortable part
The predecessor is on the Top 100 and the successor is not. For the next few days, a reader browsing AIxploria’s flagship curated list will find WeatherNext 2 — the model Google has explicitly retired from new projects — and no sign of the replacement. Combined with the badge, the directory currently points readers at the older model twice over.
Fact-Checking Every Claim on the Weather Next 3 Card
With no editorial review to audit, the checkable surface is one sentence plus the metadata. We tested all of it against Google’s documentation.
| Card claim | Google’s documentation | Verdict |
|---|---|---|
| “Hourly global forecasts” | Hourly timesteps, initialised every hour, 24 runs per day, global coverage | Correct |
| “Based on satellite imagery” | Live geostationary satellite mosaics are a direct input, alongside HRES analysis | Correct, slightly partial |
| “Up to 5 km for temperature and humidity” | 0.05° station head produces 2 m temperature and dew point | Correct and well scoped |
| “Accessible via BigQuery, Earth Engine, or Google Cloud” | All three surfaces documented in the quick start | Correct but incomplete |
| Tagged “Free” | No paid Cloud contract required — but access is allowlisted | Defensible, under-qualified |
| Filed under Research & Science | Shipping in Search, Maps and the Gemini app from day one | Misleading |
| Named “Weather Next 3” | Google’s name is “WeatherNext 3”, one word | Wrong spelling |
The naming split is not cosmetic
The directory’s own predecessor card spells it “WeatherNext 2”, one word, matching Google. The new card spells it “Weather Next 3” with a space, and the slugs diverge the same way. The consequence is measurable: searching the Top 100 page for the string “WeatherNext” returns seven hits, all of them the older model, and searching for “WeatherNext 3” returns nothing. A reader who types the vendor’s spelling will not match the directory’s newest entry.
The Biggest Omission: Weather Next 3 Already Ships in Search and Maps
If you read only the AIxploria card, you would conclude this is a cloud dataset for researchers who write SQL. That is roughly half the story, and the missing half is the larger one.
What Google actually shipped on day one
Google’s announcement states the model began powering weather experiences in Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API and Google Earth Engine on the day of launch. The consumer-facing promise is up to 50% more accurate precipitation forecasts when planning a day or more ahead.
That is the headline. A forecasting model reaching billions of people through Search and Maps is a materially different proposition from a dataset behind an access form, and none of it appears on the card. The blurb names only the three cloud surfaces, and the category places the model among academic search engines and citation tools.
Why the framing misleads
Directory readers use category and blurb as a shortlisting filter. Someone evaluating weather intelligence for logistics, agriculture or energy planning would reasonably skip a Research & Science entry describing a BigQuery dataset. They would be skipping a model already serving forecasts in products their staff use daily — and one with variables built specifically for their sector, including 100-metre wind speeds at turbine height and full solar irradiance components for generation forecasting.
What "Free" Really Means for Weather Next 3
The card marks the model Free, with a price of zero and an explicit accessible-for-free flag. Unusually for this series, that stands up better than expected — but it is not the whole picture.
The allowlist and the waiting period
Google’s quick start is unambiguous that no paid contract is needed: “Anyone with a Google Account, including researchers, students, and non-Google Cloud customers can access the data.” One approval covers Cloud Storage, BigQuery and Earth Engine together.
What the card omits is the gate. Real-time operational data is available on request, through a form, with approval “typically approved within 5–7 business days”. So Free is accurate and available now is not. For anyone treating a directory listing as a shortlist of things they can try this afternoon, a working week of lead time is the material fact.
The licensing split nobody mentions
Weather Next 3 data carries two licences depending on age. Anything relating to less than an hour ago, plus all forecasts, falls under Google’s real-time experimental data terms. Anything an hour old or more is licensed under Creative Commons Attribution 4.0, and real-time data transitions automatically once it crosses that threshold. Google also labels the whole system experimental and explicitly not a source of severe weather warnings.
Where Google's Own Numbers Disagree With Each Other
Not every inconsistency here belongs to the directory. Two of Google’s own surfaces contradict each other, and anyone quoting a headline figure should know which one they are repeating.
Fifty per cent or sixty
The developer guide states the model achieves “up to a 50% reduction in Brier score and CRPS compared to numerical weather prediction baselines when evaluated against IMERG observations”. The announcement blog, published one day earlier, claims “a Continuous Ranked Probability Score (CRPS) improvement of up to 60% against IMERG, 30% for MRMS, and 10% against rain gauge measurements for early lead times”. Same metric, same benchmark, ten points apart. A separate 50% figure in the blog refers to consumer precipitation accuracy, which is a different claim again and easy to conflate.
August or September
The developer guide’s at-a-glance table gives a release date of August 2026, and the model lineage section repeats it. The announcement blog and the arXiv paper are both dated 3 September 2026. The same pattern shows on the previous generation: the lineage lists WeatherNext 2 as June 2025 while its announcement is 17 November 2025. The most likely reading is that the lineage dates track research availability and the blog dates track product launch — but Google never says so, and the two sets of dates are presented identically.
Weather Next 3 Against the Alternatives on Its Own Card
The card lists 43 alternatives. One of them is the genuine competitor, and the directory’s own ordering places it far above.
| Attribute | WeatherNext 2 | Weather Next 3 |
|---|---|---|
| Spatial resolution | 0.25° (~25 km), all variables | 0.05° / 0.1° / 0.25° by tier |
| Temporal resolution | 6 hours | 1 hour |
| Initialisation | Every 6 hours (4 per day) | Every hour (24 per day) |
| Initialisation inputs | HRES analysis only | Live satellite mosaics + HRES |
| Ensemble members | 64 | 64 |
| Precipitation interval | 6-hour accumulation | 1-hour accumulation |
| Directory rank badge | #27 | #41 |
NVIDIA Earth-2 and the ordering problem
NVIDIA Earth-2 appears on the Weather Next 3 card’s own alternatives strip, described as a planetary digital twin simulating climate at 2 km resolution. In the Research & Science category it sits at position 5 — verified directly, on page one. Whatever ranking signal the directory uses, it currently favours a rival over both Google models.
Migration is not free either
Teams moving from the previous generation face renamed variables — height indicators moved from prefix to suffix, so 2m_temperature becomes temperature_2m — and a precipitation accumulation interval that changed from six hours to one. Any pipeline summing rainfall over multi-day windows needs its arithmetic revisited. None of this appears on either card, which is fair enough for a directory, but it is why “just switch to the newer one” is not a five-minute job.
What the Weather Next 3 Listing Means for Your Business
Two practical conclusions come out of this, and they point in different directions.
If you are evaluating weather data
Take the listing as a pointer and go straight to Google’s developer documentation. The card will not tell you about the allowlist, the 5–7 day approval, the two-tier licensing, the experimental disclaimer or the fact that the model is already in Maps and Search. Those five facts determine whether the model is usable for you, and none of them are on the page. Businesses building forecasting into operations should treat this as a data management and analytics project with a procurement step, not a sign-up.
If you use directories to shortlist tools
This case is a clean illustration of a general failure. Weather Next 3 is well described and badly placed. Every individual field on the card is defensible, yet the composite — a #41 research tool, behind its predecessor, absent from the curated lists — misrepresents a flagship release that shipped into consumer products the same week. Ranking and categorisation carry more weight with readers than blurbs do, and they are the parts directories verify least.
How to Read Any AIxploria Card After Weather Next 3
Four checks, each costing under a minute, would have surfaced everything in this article.
Follow the rank badge’s own link
Click through to the badged category and look for the tool. Three consecutive cards in this series have failed that test in three different ways: absent entirely, replaced by a predecessor, and now ranked below the model it supersedes.
Read the vendor’s launch post, not just the card
The single biggest fact about this release — Search, Maps and Gemini integration on day one — is absent from the listing and unmissable in Google’s announcement.
Check the vote count behind the stars
A 4.4-star rating sounds like a verdict. It came from seven votes. The predecessor’s 4.6 came from seven as well. Neither number carries information about quality.
Ask what “Free” is hiding
Here it hides an access form and a week’s wait. Elsewhere in this series it has hidden a card whose own FAQ answered “no” to the same question.
Frequently Asked Questions
Is Weather Next 3 the same as Google’s WeatherNext 3?
Yes. The AIxploria listing writes the name with a space; Google writes it as one word. Same model, same launch, different typography — and the directory’s own predecessor card uses Google’s spelling.
Can I use Weather Next 3 for free right now?
Free, yes; right now, no. There is no charge and no paid Google Cloud contract requirement, but real-time data needs an allowlist request that Google says is typically approved within five to seven business days.
Is the #41 rank badge wrong?
We cannot show that it is. What we verified is that the model is absent from the first 36 positions in that category, that its predecessor sits exactly where its own badge claims, and that the badge is therefore a working ordinal. The number appears to be accurate and unflattering.
How accurate is Weather Next 3 compared with the previous model?
Google reports up to a 5× improvement in resolution and, depending which Google page you read, a 50% or 60% CRPS improvement on precipitation against IMERG. Brightband’s independent Operational WeatherBench provides live third-party evaluation.
Should I switch from WeatherNext 2?
Google recommends WeatherNext 3 for all new projects and has published a deprecation guide, so for new work the answer is yes. Existing pipelines need variable renaming and a change from 6-hour to 1-hour precipitation accumulation before they will run.
References and Further Reading
Introducing WeatherNext 3, our most advanced and accurate global weather AI model
WeatherNext 3 model guide — Google for Developers
WeatherNext research and benchmarks
Quick start: Accessing WeatherNext forecasts
WeatherNext 3: Increasing resolution and performance of global weather models with raw observations
AIxploria Research & Science category
Brightband Operational WeatherBench live evaluations
WeatherNext 2: Our most advanced weather forecasting model
GPT-6 Astra Is Now Available on AIxploria
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.