What is a knowledge cutoff?

The model stopped reading on a date, and it does not know what date

A language model learns from a fixed body of text assembled before training starts. What is an LLM covers the two training stages; the point here is that both use data with an end date. Nothing published after that date is in the model. Not the election, not the price change, not the news about the model itself.

Vendors publish that date on the model card, and increasingly they publish two. Anthropic lists a "training data cutoff", the last date any data was included, and a "reliable knowledge cutoff", an earlier date before which coverage is dense enough to trust. The gap between them exists because the web takes months to write about an event: a release in March has a handful of pages in April and thousands by September. A model trained in July knows March well and July barely.

The model itself can't tell you either date. It has no calendar and no record of its own training. When asked "what is your knowledge cutoff", it produces the most likely answer for that question in its training data, which is usually a date it read about an earlier model. Anthropic's published system prompts tell Claude its own cutoff date in plain text, which is the admission that the model wouldn't know it otherwise. Take the date from the card, never from the chat.

How stale a model is on the day it ships, measured

Training takes months, and evaluation and safety work take more, so a model launches already behind. The chart reads eight model cards: the reliable or stated knowledge cutoff against the public release date, in whole months.

Two patterns. OpenAI's flagships have shipped furthest behind, ten months for both GPT-4.1 and GPT-5, while Anthropic and Google have run at two to four. And the gap only grows in service: GPT-4o was still ChatGPT's default model a year after launch, by which point its knowledge ended 19 months earlier. A model in daily use is typically a year or more behind the day you ask it something.

The dates above are the vendors' own, read from the model cards linked at the end of this page. The AI model timeline has release dates for every model the site tracks.

What the model does when you ask past the line

It answers. That's the problem. A model has no lookup step that returns "not found"; it generates the most plausible continuation, and for a question about last month that continuation is built from the patterns of similar questions in the training data. Ask about a product released after the cutoff and the model describes what such a product would plausibly be. Ask for the current price and it gives the price at cutoff, stated in the present tense.

What is an AI hallucination explains why the model is confident rather than uncertain, but the cutoff version has a tell you can use: the answer is coherent and slightly old. The model name is one version back, the price is last year's, the "latest" feature is the one from the training window. A stale answer doesn't look wrong; it looks like the past.

The fix is search, and it works on models that have none

The cutoff can't be moved, so the answer is to hand the model current text at answer time. Two mechanisms do that. Uploading a document puts your text in the prompt, and the model answers from it. Web search retrieves pages and does the same with the open web. Both are grounding, and both are the mechanism what is RAG describes: the model still predicts, but from material that postdates its training.

In Whizi, web search is on by default, sits in the composer next to the attach button, and applies to whichever model you pick. That last part is the useful bit. Claude or a Llama model: with the toggle on, each answers a question about this week from live sources. The toggle grants permission rather than issuing a command, so a request for a poem doesn't trigger a search and a question with a date in it does. A search costs no credits; the message costs the credit rung of the model that answered it, the same as without search.

One limit is honest to state. Search fixes staleness; it doesn't fix credibility. An out-of-date blog post is retrieved as readily as a vendor page. For a price or a specification, open the cited link before acting on it, and for anything that matters ask a second model with search on and compare.

Assume six months behind, then sort the question

Assume every model is at least six months behind, and a model you've used for a while is a year or more behind. Then sort the question. If the answer depends on facts about the world after that line, turn search on or upload the source. If the answer depends on reasoning about something already in the conversation, the cutoff doesn't matter and search can be off.

Check the model card, not the model, for the date. And when a model volunteers "as of my last update", read it as the model quoting its training data rather than reporting a fact about itself, and treat everything after that phrase with the same suspicion.