Knowledge cutoffs — what the model has not seen — The Prompt Engineering Path — Foundations — Nerd Level Tech

Every language model is trained on a fixed snapshot of the internet. The day that snapshot was taken is the knowledge cutoff. After that date, the model knows nothing — it does not silently update, it does not "stay current", and it cannot tell you what happened last week.

Two details make this harder than "one date per model".

The horizon is two dates, not one. Anthropic publishes both for every model: a training data cutoff (the outer edge of the data that went in) and a reliable knowledge cutoff (the date through which the model's knowledge is actually dependable). They can sit months apart, and it is the reliable one that should govern what you trust. Look yours up on the model comparison table — every vendor publishes the equivalent, and every model has its own pair.

The model does not reliably know either date. This is the part that catches people, and the next section is a live demonstration of it.

The principle survives all of it: the model has a horizon, and beyond that horizon, every fact it produces is a guess shaped by its training data, not retrieved from reality.

Module 8's four lessons in order:

What this looks like in practice

We asked the model a question whose answer depends on events well past its cutoff:

Who won the most recent Egyptian Premier League title?

A poorly-tuned model would invent a confident answer. Here is what we got:

I don't have access to real-time information, so I can't tell you who won the most recent Egyptian Premier League title. My knowledge was last updated in April 2024, and league seasons continue beyond that date.

To find the current champion, I'd recommend:
- Checking the Egyptian Premier League's official website
- Looking at recent sports news from reliable sources
- Visiting sports websites like ESPN, BBC Sport, or specialized football sites

Could you let me know what year or season you're asking about? I might be able to help if it's from before my knowledge cutoff date.

Captured from Claude Sonnet 4.5 (claude-sonnet-4-5) on 2026-04-27. Re-runs may differ slightly.

The model did two things right and one thing you must not copy.

Right: it named the gap explicitly ("I don't have access to real-time information") instead of inventing a champion, and it pointed to authoritative sources the user could actually check.

Wrong: the date. The model volunteered "my knowledge was last updated in April 2024". Anthropic's own model table puts Claude Sonnet 4.5's reliable knowledge cutoff at January 2025 and its training data cutoff at July 2025 — so the model understated its own horizon by nine months on one measure and fifteen on the other.

Sit with that, because it is the whole module in one example. The reply looks like the gold-standard hedge. It has the shape of a careful, self-aware answer. And the single most checkable fact in it is wrong — in a response about the limits of its own knowledge.

A model does not have privileged access to facts about itself. Its own cutoff is not introspection; it is one more claim generated from training data, and a poorly-attested one, because a model is trained largely on text written before it existed. This is the failure mode that survives every instinct you have, since the answer arrives wearing exactly the humility you were hoping to see.

So: take the behaviour from this reply and never take the date. The behaviour you want a production assistant to default to is naming the gap and handing over a recovery path. The date comes from the vendor's published table, which is a page that gets corrected — not from asking the model.

Try it yourself. Ask any assistant "what is your knowledge cutoff?", then check its answer against its vendor's model page. Doing this once, with your own eyes, is worth more than the rest of this lesson.

Two ways to find a model's horizon

What we did above

Ask the model

Where the answer comes fromGenerated from training data

Accuracy in our captureOff by 9–15 months

Gets corrected when wrongNever

Pros

Zero effort
Available inside the conversation you are already having

Cons

It is a guess, not introspection
The model was trained mostly on text written before it existed
Arrives wearing exactly the humility you were hoping to see

About fifteen seconds

Read the vendor's model page

Where the answer comes fromPublished by whoever trained it

AccuracyAuthoritative

Gets corrected when wrongYes — it is a maintained page

Pros

Separates reliable-knowledge from training-data cutoff
Same page carries the model's retirement date
Works the same way for every vendor

Cons

You have to remember to go and look
Each vendor words and dates it differently
Tells you the horizon, not whether your specific topic was well covered

Why this matters for Bayt Coffee

The Bayt Coffee assistant Hagar built in Module 7 has the same blind spot. If a customer asks "did you win an award at this year's specialty coffee expo?", the model has no way of knowing — but its default training pushes it to sound helpful, which means making something up.

The next three lessons fix this in three layers: hedge language for general unknowns (this lesson sets the bar), the "I don't see that in the document" trigger for grounded RAG-style replies, and a tour of the most common hallucination patterns to test against before shipping.

Next: hedge language as a system-prompt instruction. :::