What is ChatGPT's knowledge cutoff, and why does it vary?

A knowledge cutoff is the date after which a model saw no training data. Anything that happened later is invisible to the model unless it fetches a web page at answer time.

OpenAI's model documentation, as verified on August 22, 2026, lists these dates:

OpenAI describes chat-latest as the latest Instant model currently used in ChatGPT. Its cutoff sits about five and a half months before the GPT-5.6 family's. The same docs warn that the snapshot "will be regularly updated", so these dates can move. Newer models, such as the gpt-6-astra API model, may carry different cutoffs that our evidence base does not record. Check OpenAI's model pages before relying on any date.

"ChatGPT" is a product, not one model. It routes requests between a reasoning family and a faster Instant path. Each path has its own training date.

Plan tier changes the picture too. Secondary reports, not confirmed by OpenAI, say Free and Go users get GPT-5.6 Luna, Plus users get GPT-5.6 Sol, and higher tiers get a GPT-6 Pro model. Two people asking the same question on different plans can therefore get answers built on different knowledge.

What this means for you: a single test of "what does ChatGPT know about us" is a sample from one route. It is not a reading of ChatGPT as a whole. Any test that does not control for routing and search state measures an unknown mixture.

The cutoff only governs answers ChatGPT gives from memory. When ChatGPT searches the web, it can read pages published yesterday. OpenAI's help docs say ChatGPT search typically rewrites your question into one or more targeted queries and sends them to search providers.

The catch is that searching is the model's choice, made per query. Published estimates of how often ChatGPT searches range from about 18% (Profound) to 80.5% (cloro), with Semrush reporting 34.5% of US queries in February 2026. They measure different things on different samples, so treat none of them as the true rate.

Two more conditions shape what you see:

• Memory. OpenAI says that if memory is enabled, ChatGPT may use saved memories when rewriting a search query. Logged-in tests are not reproducible.
• Citations. If an answer shows no sources, you can't tell whether it came from search or from training data.

Our guide to search-on versus search-off answers covers how the two modes produce different claims about the same brand.

Say you run a hypothetical SaaS company that changed its pricing in March 2026. That change sits after every cutoff in the table above. No ChatGPT model in OpenAI's docs could have learned it in training. On the Instant path, anything after August 31, 2025 is missing too, so a rename in December 2025 would also be invisible from memory.

In practice this produces three patterns:

1. Search-off answers state the old fact. The model repeats the last version it saw, often with confidence.
2. Search-on answers can state the new fact, but only if ChatGPT retrieves a page that says it.
3. Answers disagree across sessions, because routing and the search decision vary.

The same logic applies if your company launched recently. Our page on whether ChatGPT can know a new company covers that case. Nothing you do this quarter changes a frozen model. A future model with a later cutoff might learn the new fact, but OpenAI publishes no schedule, and how long AI takes to update has no fixed answer.

You can't move a cutoff. You can make sure that when ChatGPT does search, it can find and read the current facts.

1. List your changes against the dates. Note any pricing, product, ownership or leadership change after August 31, 2025 and after February 16, 2026. Those are the facts at risk in search-off answers. The free knowledge cutoff check walks through this.
2. Check per-bot robots.txt rules. OpenAI's crawler page says OAI-SearchBot controls ChatGPT search, and sites that block it "will not be shown in ChatGPT search answers". GPTBot governs training only. OpenAI says each setting is independent.
3. Check the firewall. OpenAI's help docs say eligibility also requires your host or CDN to allow its published searchbot IP addresses. A robots.txt allow means nothing behind a 403. The AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.
4. Put the facts in the HTML. Vercel's crawler study found ChatGPT's crawlers fetch JavaScript but don't execute it. A price that only appears after scripts run may be invisible to them.
5. State each fact plainly on one page. A clear sentence ("Our Pro plan costs $X a month") gives the engine something it can quote.

For a full sequence, see the playbook on correcting outdated facts past a cutoff. To see where you stand today, the free checker on this site shows whether AI engines name you for up to three buyer questions.