Knowledge cutoff

Data Beginner Also called: training cutoff, training data cutoff

Definition

The date a model's training data stops. It knows nothing newer unless it searches the web or you tell it. Ask about last week without search on and you get a guess or an admission. Model makers publish the date; Anthropic also publishes an earlier "reliable" date, because knowledge of the final months before the cutoff is thinner.

In plain English

The model went to school until a certain date and then stopped reading the news. Ask about anything after graduation and it either admits it doesn’t know or, worse, guesses with a straight face.¹

In the real world

“I don’t have information past [date]” in a chat reply.
A confident summary of a product that was discontinued last year.
The model assuming it’s still the year it was trained.

When to use it

Before asking about anything recent: prices, versions, news, people’s jobs.
Comparing models. Anthropic publishes two dates per model, the training cutoff and the earlier reliable one.¹

When not to use it

Don’t treat it as a hard wall. Models have patchy knowledge of the months right before the cutoff, which is why the “reliable” date exists.¹ When freshness matters, search or paste.

Good, bad, ugly

Good: a known limit. You can plan around a date.
Bad: the model doesn’t always know what it doesn’t know.
Ugly: stale facts delivered in the same confident tone as fresh ones. See hallucination.

Key facts

Two dates: Training data cutoff, and the earlier reliable knowledge cutoff
Workaround: Web search, or paste the current information into the chat
Symptom: Confident answers about events that happened after the cutoff
Not the same as: A hallucination. Outdated is a date problem. Invented is a guessing problem.

Common questions

Why doesn't the AI know about something recent?

Its training stopped at a date. Anything after that isn't in the model. If the tool has web search, turn it on. If not, paste what it needs to know.

What's a reliable knowledge cutoff?

Anthropic's term for the date up to which the model's knowledge is solid. Data from the last months before the training cutoff is thinner, so the reliable date is earlier.

Does the model know today's date?

Only if the app tells it. On its own, it may assume it's still the cutoff year. Say the date if it matters.

References

Models overview: reliable knowledge cutoff and training data cutoffAnthropic, Claude Platform DocsView source