Every AI chatbot has a knowledge cutoff: the date after which its training data stops, so it has no built-in memory of anything that happened later. Ask it about something that broke the week after that date, and it will either say it doesn’t know or, worse, guess. Understanding where that line falls — and why it exists at all — explains a lot of AI’s most confusing behavior.
What a knowledge cutoff actually is
A knowledge cutoff is the point in time beyond which a model has not been trained on new data. Large language models like the ones behind ChatGPT and Claude don’t browse the internet by default — they learn patterns from a fixed snapshot of text gathered before training began, then that snapshot is frozen. Everything the model “knows” comes from that snapshot, the way a printed encyclopedia only reflects the world as of its publication date.
This is different from a model simply being old. A brand-new large language model released today can still have a cutoff months in the past, because collecting, cleaning, and training on data takes time — then testing the model for safety and accuracy takes more time before it ever reaches users.
“Reliable” cutoff vs. training cutoff
Cutoffs aren’t always one clean date. Anthropic, for instance, documents two separate figures for each Claude model: a “training data cutoff” (the outer boundary of data used in training) and a “reliable knowledge cutoff” (the date through which the model’s knowledge is most extensive and accurate). As of August 2026, Anthropic’s own model documentation lists Claude Opus 5’s reliable cutoff as May 2026, while Claude Sonnet 5 and Claude Fable 5 sit at January 2026 — even though all three are current models. OpenAI’s GPT-5.6 models, released in mid-2026, cite a February 2026 cutoff. The gap between a model’s release date and its cutoff date is normal, not a bug: it’s the lag between finishing training and shipping a product.
Because every new model generation pushes this date forward, treat any specific cutoff as a snapshot in time rather than a permanent fact — always check a model’s own documentation for its current figure.
How AI works around its own cutoff
Rather than retraining a model every day, most AI products now pair it with live lookup tools. When a chatbot needs information past its cutoff, it can run a web search and read the results before answering — a pattern known as retrieval-augmented generation. Claude and ChatGPT both support this: enabling web search in Claude lets it fetch current pages and cite its sources instead of relying purely on memorized training data.
This fixes the “doesn’t know” problem but not every problem. A model without search enabled — or one that misjudges whether a search is needed — can still answer a post-cutoff question from memory and get it wrong with total confidence, a failure mode known as hallucination. Search reduces this risk; it doesn’t eliminate it, since the model still has to judge which retrieved pages are accurate.
Why it matters
The practical takeaway is simple: don’t assume a chatbot’s answer about anything recent is current unless you know it actually looked it up. If a tool doesn’t cite sources or show it searched, treat time-sensitive claims — prices, product versions, who holds what job, breaking news — as possibly stale. For programming, this also matters concretely: a model trained before a library’s latest release may recommend outdated syntax or a deprecated API, even while sounding completely certain.
Knowledge cutoffs aren’t a flaw to be embarrassed about; they’re a direct consequence of how these models are built. Knowing a model’s cutoff — and whether it can search past it — is the single fastest way to judge how much to trust what it tells you.