Products, laws and numbers here change fast. Checked October 2026.
In plain words
A language model learns by reading a huge pile of text. At some point the builders stop adding to the pile and start training. That date is the knowledge cutoff. Anything that happened later, a new phone, an election result, last week's match, simply is not in the model.
Training and testing take months, so a model usually reaches you well after its cutoff. Ask it about something recent and it may say it doesn't know, or worse, guess from older facts and sound sure.
Many chat apps now get around this by searching the web before answering. Then the reply can be current, but only because the app fetched fresh pages, not because the model itself learned anything new.
Everyday example
A friend back from a long trek
A friend returns from three months trekking with no phone signal. Ask who won the series and they will tell you what they last heard, maybe with confidence. Hand them today's newspaper and they can answer correctly. The trek start date is the cutoff. The newspaper is the web search.
Where the example breaks: Your friend knows exactly when they left. A model often does not know its own cutoff precisely, and its knowledge thins out gradually in the last months before it rather than stopping on one clean day.
How it actually works
The cutoff is a property of the training data, not of the app you use.
Freeze the data
Builders collect web pages, books and other training data up to a chosen date, then clean it and remove duplicates. Nothing written after that date goes in.
Train, test, release
Pre-training and safety testing take months. Anthropic's Transparency Hub lists one model with a January 2026 cutoff released in June 2026. Labs publish each model's cutoff in their developer docs.
A blurry edge
Less has been written about the final months before a cutoff, so the model knows them thinly. Researchers at Johns Hopkins found a model's "effective cutoff" often differs from the date its makers report, and can differ by topic.
Bridging the gap
Apps add fresh information at question time: a web search, your uploaded files, or a company's documents through RAG. The model reads that text in its context window and answers from it.
- Also called
- Training cutoff, training data cutoff, cutoff date
- Typical gap
- Several months between cutoff and release. Anthropic lists one model with a January 2026 cutoff that launched in June 2026
- As of October 2026
- The newest flagship models from OpenAI and Anthropic list cutoffs in the first half of 2026
- Research finding
- "effective cutoffs often differ from reported cutoffs" (Cheng and colleagues, Johns Hopkins, 2024)
A newer cutoff makes a model more current, but a model that can search is current in a different way: it reads today's pages.
Questions people ask
What does knowledge cutoff mean in ChatGPT?
It is the date the training data for the model inside ChatGPT ends. The model itself knows nothing after it. When ChatGPT searches the web, it can still answer about recent events, because it reads fresh pages and summarises them for you.
Why does the AI give me old information?
Because its training data stops at the cutoff, and some months pass before release. If the app did not search the web for your question, the model answers from what it learned, which may be out of date. Ask it to search, or check the date on its sources.
Can I ask an AI what its knowledge cutoff is?
You can, but the answer may be wrong. Models often do not know their own cutoff precisely. The reliable place to look is the company's model documentation, which lists the cutoff for each model.
Does web search remove the knowledge cutoff?
Not exactly. The model's built-in knowledge still stops at the cutoff. Web search adds current pages to the prompt so the model can read them. The answer is only as current and correct as the pages it found.