What is the knowledge cutoff of an AI model?

The knowledge cutoff is the date after which a language model saw no more data during training: whatever happened later simply does not exist for the model. An LLM learns by reading a huge collection of text, and that collection is closed at some point, usually months before the model reaches the public. From then on its knowledge stays frozen, like a printed encyclopedia. A law that was just amended, a new price list, a supplier's new CEO: none of it ever reached the model. And the model often does not know what it does not know: it answers with the old version, as confident as ever. Vendors state the date in their model documentation, and anyone using an AI assistant on recent topics also needs to know whether the system can look up current sources.

See also: Grounding · Anchoring a model's answers to verifiable sources instead of its memory: RAG is the main technique used to achieve it.

Stated date and effective date

The cutoff exists because collecting and cleaning the data, then training the model, takes months, and retraining is expensive. That is why the date always trails the release. OpenAI and Google publish a single "knowledge cutoff" field for each model. Anthropic publishes two: the training data cutoff, which is where the collection ends, and a reliable knowledge cutoff, the same or earlier, through which the model's knowledge is most extensive and reliable. Anthropic does not say why; a plausible explanation is that the web holds little about the last months before the collection closes, because analysis and commentary arrive afterward. A 2024 study by Cheng and colleagues found that the effective cutoff of many models differs from the stated one and varies by topic, partly because newer web crawls contain much old material that deduplication fails to remove. When a vendor states two dates, it is therefore safer to rely on the reliable knowledge one.

Parametric and contextual knowledge

What the model learned in training is called parametric knowledge: it is written into its weights, stops at the cutoff and carries no record of where it came from. Contextual knowledge is the text the system places in front of the model at question time (a document, a search result, a record from a business system), inside its context window. It is as current as the source it comes from. RAG and web search exist precisely to base the answer on contextual rather than parametric knowledge, which is what grounding means. The other forms of memory an agent can have are covered in the entry on agent memory types.

An enterprise example

A company gives its sales team an assistant that answers questions about tax incentives for customers. A rate changes after the model's cutoff: the assistant keeps quoting the old one without hesitation, and a quote goes out wrong. It is not a hallucination in the strict sense, since the figure was true until shortly before, but the effect is the same. Switching to a newer model only pushes the problem forward. What solves it is having the assistant read the current official text and showing the user the date of the source it used.

Why it matters for decision makers

When I evaluate an AI assistant for a client I ask three questions.

• What is the cutoff of the model in use?
• Which current sources can it consult, and how often are they refreshed?
• What changes when the vendor swaps the model?

Then I run a test: I ask the assistant about something recent and check whether it cites a dated source. The dates change with every release, so they should be read in the model's official documentation at the time of the choice. Asking the model itself is not enough: its answer comes from the same parametric knowledge you are trying to check.