It Does Not Know What Happened After the Cutoff
An LLM learns how words connect from a large amount of collected text. The collection of training text stops at some point. That point is the knowledge cutoff. What the model knows comes from text up to this date.
As a result, the model does not know about events after the cutoff, or about prices and rules that changed later. Instead of saying it does not know, it may answer with outdated content as of the cutoff. Anthropic also explains that a model's knowledge is most extensive and reliable on information and events up to the cutoff.
The Cutoff Is Set by Version
The cutoff is set for each model version. Even among models from the same company, the cutoff may be the same or may differ. Anthropic and OpenAI list the cutoff for each model in their model lists. You can check the cutoff of the model you use in the developer's list.
Anthropic lists two dates separately: a reliable knowledge cutoff and a training data cutoff.
How to Fill In Newer Information
When you need information from after the cutoff, pass it to the model from outside. One way is to have the model read results found with a web search feature and answer from them. The other is RAG, which finds the relevant parts of your own documents and passes them along with the question.
When neither is available, check the date-dependent parts of the answer yourself. When choosing a model, look at whether it has features for passing in the information you need rather than at how recent its cutoff is. The Tool use condition on the pages for choosing by use marks whether a model can call outside features such as search.