Knowledge Cutoff

Knowledge cutoff is the date until which the model saw the training data. After it, the model knows nothing: no new laws, no updates to platform algorithms, no fresh research. GPT-4o has a cutoff in April 2024, Claude Sonnet 4.6 has a cutoff in August 2025.

In my practice, this is the first thing I check when using LLM for tasks that are sensitive to relevance: analysis of social network algorithms, tax legislation, conditions of advertising platforms. If a task requires up-to-date data, I either connect web search (Perplexity, ChatGPT with search) or manually insert the current document into the context.

The problem is not the cutoff itself - the problem is that the models do not warn the user when they respond based on outdated data. GPT-4 will confidently tell you about the “current” conditions of Yandex Direct, based on data from two years ago. The rule that I follow: for any fact that could have changed over the year, verification through a current source is required.