ML//knowledge cutoff

The knowledge cutoff is the date at which an LLM's training data was frozen. Everything published after it is invisible to the model unless it is retrieved in real time; extended thinking does not change that, since a model can reason deeply about what it learned before the cutoff but cannot think its way to later facts.

The knowledge cutoff is the date at which an LLM's training data was frozen. Everything published after it is invisible to the model unless it is retrieved in real time; extended thinking does not change that, since a model can reason deeply about what it learned before the cutoff but cannot think its way to later facts.

It is not a clean cut: training data is a snapshot of the web at crawl time, processed and filtered over months. The cutoff is the latest date in the corpus rather than a single collection event.

For web developers, even with perfect AEO, a site's content only enters a model's knowledge if it was crawled before the cutoff and survived filtering. Getting into training data is a pipeline that takes months.

Real-time retrieval bypasses the cutoff: Perplexity, ChatGPT with search and Google AI Overviews fetch live content. This is faster but less sticky, because the model does not know the site; it just found it.

Training data and retrieval give two kinds of knowledge. Training data shapes the model's priors (it believes things), while retrieved content is contextual (it reads things). Being in the training data means the model recommends a site unprompted; being in retrieval means it cites the site when asked.

As of early 2026 the cutoffs were around April 2024 for GPT-4 and for Claude 3.5, and varied for Gemini. Each new model generation pushes the cutoff forward.