Before the training horizon
09 Jan 2024
OpenAI launches GPT Store
Part of the model world
verified in 8 of 9 models without web access
An early product anchor most models still cover.
Grounding Page Results
The knowledge gap of AI models
Which world events, products and software states does an AI model still know without live web access?
Headline finding
Without grounding, AI models often answer current questions from a world that is 1 to 2 years in the past.
New products, roles, drug approvals and events only become reliably visible once they are grounded via current sources.
Six examples show what reality is missing without grounding.
From new roles via sport events to product launches and drug approvals: anything that happens after a model’s training horizon is not reliably available without grounding.
Before the training horizon
14 Jul 2024
Spain wins UEFA Euro 2024
Mostly verified
verified in 6 of 9 models without web access
A sport anchor close to many model horizons.
After most training horizons
08 May 2025
New pope elected
Knowledge edge
verified in only 1 of 9 models without web access
verified: GPT-5.5
A change at the head of a world institution — barely any model knows it.
After most training horizons
27 Jul 2025
England wins Women’s Euro 2025
Knowledge edge
verified in only 1 of 9 models without web access
most recent verified hit · GPT-5.5
Reliable only with grounding
09 Sep 2025
Apple introduces iPhone 17
Missing without grounding
verified in 0 of 9 models without web access
Public reality, but not yet model reality.
Medical knowledge gap
01 Apr 2026
FDA approves Foundayo for obesity
Missing without grounding
verified in 0 of 9 models without web access
Newly approved medication for weight reduction in adults with obesity or corresponding overweight.
The point
New reality emerges faster than model knowledge is updated.
That is why products, medications, role changes, events and organisations need current, structured and citable sources.
Tested model horizons
The tested models show different knowledge horizons. Most families end somewhere in 2024; individual hits reach further.
-
GPT-5.5 June – July 2025
-
-
-
-
-
Gemini 3.1 Pro May – July 2024
-
-
-
Single-run tests fluctuate by 2–3 months, so we report family windows instead of pinpoint dates.
A model can sound current and still not know what happened after its cutoff.
Without grounding, model knowledge has a visible edge.
The question is not how smart the model sounds. The question is where its world ends.
Evidence
Tested model horizons
Reference data behind the page. Estimated horizon per model, verified hits and details — not a ranking.
| Rank | Model | Provider | Estimated horizon | Verified | ||||||||||||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| #1 | OpenAI | June – July 2025 | 10 / 10 | |||||||||||||||||||||||||||||||||||
Top-10 verified statements · most recent first
|
||||||||||||||||||||||||||||||||||||||
| #2 | OpenAI | May – July 2024 | 10 / 10 | |||||||||||||||||||||||||||||||||||
Top-10 verified statements · most recent first
|
||||||||||||||||||||||||||||||||||||||
| #3 | OpenAI | May – July 2024 | 10 / 10 | |||||||||||||||||||||||||||||||||||
Top-10 verified statements · most recent first
|
||||||||||||||||||||||||||||||||||||||
| #4 | OpenAI | May – July 2024 | 10 / 10 | |||||||||||||||||||||||||||||||||||
Top-10 verified statements · most recent first
|
||||||||||||||||||||||||||||||||||||||
| #5 | OpenAI | May – July 2024 | 10 / 10 | |||||||||||||||||||||||||||||||||||
Top-10 verified statements · most recent first
|
||||||||||||||||||||||||||||||||||||||
| #6 | Google Gemini | May – July 2024 | 10 / 10 | |||||||||||||||||||||||||||||||||||
Top-10 verified statements · most recent first
|
||||||||||||||||||||||||||||||||||||||
| #7 | Google Gemini | May – July 2024 | 10 / 10 | |||||||||||||||||||||||||||||||||||
Top-10 verified statements · most recent first
|
||||||||||||||||||||||||||||||||||||||
| #8 | Google Gemini | April – June 2024 | 10 / 10 | |||||||||||||||||||||||||||||||||||
Top-10 verified statements · most recent first
|
||||||||||||||||||||||||||||||||||||||
| #9 | Google Gemini | April – June 2024 | 10 / 10 | |||||||||||||||||||||||||||||||||||
Top-10 verified statements · most recent first
|
||||||||||||||||||||||||||||||||||||||
Methodology
01
Test principle
For each model, we ask one question per event. Models answer without live web search, grounding, or tools.
02
Verification
Correct answers are externally verified. The most recent correctly-answered event defines the model’s knowledge horizon.
03
Family buckets
Single-run results swing by 2–3 months, so we report family windows. Suspicious cutoffs are re-tested at least 5 times.
04
Repeat cadence
Model knowledge is frozen at pretraining. The comparison is repeated weekly; new model versions are added with the next run.
Why these events?
Reveal-at-Date rule
All examples are Reveal-at-Date events: the concrete answer could not have been known before the event date and is unambiguously verifiable afterwards.
Qualify
Election winners, final winners, newly elected popes, unveiled products, software releases, regulatory approvals, court rulings.
Do not qualify
Scheduled conferences, trade fair dates, announced sport events, Olympic openings.
What this has to do with Grounding Pages
Grounding Pages do not close the training itself. They close the access gap.
They make new or changed entities findable, understandable and reliably usable through current, structured and citable sources — exactly what AI systems need when retrieval, live web search or other grounding mechanisms kick in.
- new products
- new medications and approvals
- new car models
- new software states
- role changes
- location changes
- new standards
- changed organisational structures