Knowledge cutoff: definition and why it matters

A knowledge cutoff, also called a training cutoff or model cutoff date, is the point in time after which a model received no training data. Everything the model learned in training sits before that date. It forms the model's parametric knowledge, the facts it can state with no web access.

The cutoff is a ceiling, not a setting. Vendors publish it per model version, and it only moves when a new model ships. If your company launched, rebranded or changed pricing after a model's cutoff, that model cannot know it from training. For the brand-level consequences in depth, read our guide on why AI knows your old news.

People talk about "ChatGPT's knowledge" as one thing. It isn't. Two variables decide which facts you get.

Which model answers

OpenAI's model documentation, as verified in August 2026, listed a February 16, 2026 knowledge cutoff for the GPT-5.6 models (Luna, Sol and Terra). The same documentation listed August 31, 2025 for chat-latest, the Instant model path used in ChatGPT. That is roughly five and a half months of difference inside one product. OpenAI also warns these snapshots are regularly updated, so check the current figure before relying on it. Our answer on ChatGPT's knowledge cutoff covers why it varies.

Whether the answer searched

Retrieval is per query and discretionary. As of October 2026:

• ChatGPT decides when to search, and OpenAI says it rewrites your question into one or more targeted queries.
• Claude.ai search is a user toggle with no published default.
• The Gemini API's grounding generates and runs its own search queries.

When search runs, fetched pages enter the model's context window and the answer can reflect facts newer than the cutoff. When it doesn't, the cutoff is the hard limit.

AI answers about your brand come from two channels. Retrieval-off answers rely on training data alone. Retrieval-on answers can also use pages fetched for that query. Our guide on search on vs search off shows how far the two can drift.

Be plain about control. No one can change a model's training data this quarter. A retrieval-off answer will keep stating whatever the model learned before its cutoff until a newer model replaces it.

What you can control is access for retrieval-on answers:

1. List your recent brand changes and compare their dates against each target model's published cutoff. Anything newer is unreachable in retrieval-off answers.
2. Put the new facts (price, product name, launch date) in your page's HTML before JavaScript runs. Most AI crawlers other than Googlebot do not execute JavaScript, according to Vercel's crawler study.
3. Check that your robots.txt allows the search bots, such as OAI-SearchBot for ChatGPT search, and that your CDN or firewall doesn't block them.

None of this guarantees a citation. It removes the blocks that stop a current fact from reaching the answer. For a step-by-step approach, see the playbook on correcting outdated facts past a cutoff.

• "The app has one cutoff." It doesn't. Routing between models changes the cutoff, as OpenAI's own documentation shows. That's why switching AI models can change what it says about your brand.
• "Allowing GPTBot puts my news into ChatGPT." OpenAI says GPTBot governs training use only. Any training effect would land a model generation later, and nobody has measured it. Search inclusion runs through OAI-SearchBot, and OpenAI says each setting is independent.
• "A newer cutoff means the model knows my brand." A cutoff says when training data stops, not what it contains. Well-known brands appear far more often in AI answers. A small brand may be thin in training data even before the cutoff.
• "Updating my site fixes the old answer." It can affect retrieval-on answers once crawlers fetch the page. It does nothing to the frozen training data.

• Parametric knowledge: what a model knows from training, bounded by its cutoff.
• Context window: where retrieved pages go when an answer searches.
• Retrieval: fetching live pages at answer time, which is how an answer gets past the cutoff.
• Grounding: tying an answer to retrieved sources rather than training data alone.

Where can I find a model's knowledge cutoff? Vendors publish it on their model documentation pages, usually per model version. OpenAI lists a cutoff date for each API model, for example. Check the date each time you test, because vendors update snapshots and change which model a consumer app uses. The model an app serves to free users may differ from the one paid plans get, and vendors do not always say which is the default.

Why does ChatGPT give different facts about my company on different days? Two things change between answers. The question may route to a model with a different cutoff, and ChatGPT decides per query whether to search the web. An answer from training data alone reflects the cutoff. A searched answer can use your current pages. If memory is on, saved memories can also shape how ChatGPT rewrites the search query. Test logged out, with memory off, for a cleaner comparison.

Can I get my new facts into a model before its next release? Not into its training data. That is fixed at the cutoff. You can make the facts reachable for searched answers: state them plainly in server-rendered HTML, allow the search crawlers in robots.txt, and confirm your firewall lets them through. That gives retrieval-on answers a current sentence to use, though no engine promises to use it.

Is knowledge cutoff the same as the model's release date? No. The cutoff is when training data ends. The release date is when the model becomes available, which is later. A model can be in use for months while its knowledge stays fixed at the earlier cutoff date. So the newest model in an app can still miss events from the weeks or months before you used it.