What is a knowledge cutoff?

You ask a chatbot which phone is the newest, or who won last season's championship. It answers straight away, clearly and confidently. And the answer is a year old. Nothing in its tone warns you. Nobody set out to mislead you; the model just hasn't heard the news. The reason is a date most people never see: the model's knowledge cutoff.

An artificial intelligence (AI) model learns from a huge collection of text, gathered up to a certain point. Then the gathering stops. Training starts, then testing, and the model reaches the public later, often months later. The point where the collection ends is the knowledge cutoff. Events after that date were never in its reading. As one example, Anthropic lists the training data for its Claude Haiku 4.5 model as running to July 2025. The model came out in October 2025. So it started life about three months behind, before anyone had typed a single question into it. Its successor, Claude Haiku 5.5, is expected around mid-October 2026, with its own, later cutoff.¹²⁶

The cutoff is less of a clean line than it sounds. In 2024, researchers at Johns Hopkins University found that the point where a model's knowledge is most concentrated often differs from the cutoff its makers report, sometimes by a lot. Part of the reason: even a fresh sweep of the web is full of older pages, so the newest material is a thin slice of it. Anthropic's model table shows a gap like this. Next to the training data cutoff, it lists an earlier "reliable knowledge cutoff," the date through which the model's knowledge is most complete. For Haiku 4.5, those are July 2025 and February 2025. The edge is blurry.¹⁵

The tricky part: a model has no clock and no news feed of its own. When you ask for "the latest," it gives you the latest thing in its reading, which may be long out of date. Ask about a price or a sports record, and it may hand you the last figure it read, with no hint of its age. It may not be sure when its own reading stopped, either, so it can't always warn you. That's why a stale answer can sound every bit as sure as a fresh one.

Many chat apps can now look things up. That helps. OpenAI added web search to ChatGPT in 2024. It can choose to search when a question is about something recent, like news, sports scores, or stock prices, or you can ask it to. The app fetches fresh pages and hands them to the model, and the answer comes back with links to its sources. That's a different job from remembering. The model is reading something new, right then. Search has limits of its own, though. The pages it finds can be wrong too, which is why the links are there for you to open.³⁴

Three habits keep you on solid ground.

• When the date matters, ask "as of when?"
• For anything recent, turn on search, or ask the app to look it up.
• Check anything time-sensitive, like prices, rules, or schedules, against a current source.

Think back over the last few things you asked a chatbot. Did any of them have a date hiding inside?