ChatGPT knowledge cutoff dates and what web search changesVerified dates for ChatGPT, Gemini, Claude and Perplexity, and what moves them

The ChatGPT knowledge cutoff is the date the model’s training data stops, and it depends on which model answers: GPT-6 Sol, April 20, 2026; GPT-6 Luna, May 18, 2026; the GPT-5.6 models that free plans still use, February 16, 2026.

The date moves every time OpenAI ships a new model behind the same product, so any single answer goes stale. And when ChatGPT searches the web, the cutoff stops mattering for that answer, because it’s built from the pages the searches return, not from training data. That split, memory versus retrieval, is where an AI SEO platform starts: it records whether the assistant searched before it answered, since that decides where your brand can come from at all.

Gemini, Claude and Perplexity carry their own dates, and what the cutoff means for your brand is mostly whether the assistant searched or answered from memory.

What a knowledge cutoff is

A knowledge cutoff is the date a model’s training data stops, and everything after it never made it into training. The model can still talk about later events, but when it does, it’s filling gaps, not remembering.

OpenAI publishes a “Knowledge cutoff” line on every model page in its platform docs. Anthropic’s docs are more careful: Claude Opus 5.5 carries both a “Training data cutoff” and a “Reliable knowledge cutoff”, and both read June 2026. Google states a knowledge cutoff in the model cards it publishes for each Gemini release.

The part that trips people up is that the cutoff belongs to the model, not the product. ChatGPT, Gemini and Claude are products that swap models in and out, so there is no one product-level date to quote. Ask what ChatGPT’s cutoff is and the honest answer is a table, which is basically why the search results for this question disagree with each other.

ChatGPT knowledge cutoff dates by model

So ChatGPT’s knowledge cutoff date is whichever model answers your question on the day you ask, and that depends on your plan and your thinking level. Currently, GPT-6 is rolling out to eligible Pro, Plus, Enterprise and Business users: GPT-6 Sol powers Instant through Extra High, GPT-6 Pro runs on GPT-6 Astra, and GPT-5.6 Luna is becoming the default for Free and Go users.

So two people asking the same question on the same day can get answers from models whose training data stops months apart, and one of them is a generation behind the other. OpenAI shifts this routing often, so treat the table as a snapshot, not a constant.

Gemini, Claude and Perplexity cutoff dates

Google publishes months, not days, so Gemini’s knowledge cutoff date can’t be pinned sharper than March or January. The Gemini 3.8 Flash card also carries a caveat worth reading: in some domains its knowledge is limited to January 2025, in line with the Gemini 3 family. And Anthropic’s Transparency Hub gives the older Claude 5 models their dates, May 2026 for Opus 5 and January 2026 for Sonnet 5 and Fable 5.

Perplexity is the odd one out. Its docs don’t publish a knowledge cutoff for the Sonar models anywhere, and its search is always on, so what you get back is built from pages it has just read.

Its Agent API docs say the API selects across the OpenAI, Anthropic, Google, xAI, Z.AI, Moonshot AI and NVIDIA model families in addition to Sonar, though that describes the API, not the consumer app. The plain reading is that Perplexity answers retrieval-first, which is the direction all four assistants are moving.

Why the dates keep moving

ChatGPT’s knowledge cutoff date changes because the cutoff belongs to the model, and the model behind the product name keeps changing. A new model ships with a new cutoff, the product keeps its name, and your plan and thinking level pick which one answers. Models also roll out gradually, so the same chat can change models without you touching anything.

The community threads that rank for this question show the confusion this causes: people report the cutoff jumping between dates and ask whether an update moved it. Nothing moved. They were talking to a different model, and the number they remembered was that model’s.

If you publish these dates, the working rule is that a number without its source and its check date is basically already wrong. Every date here links the provider page it came from, read on the day it was checked. That’s the only honest way to publish a number that moves.

Why web search makes the cutoff matter less

When ChatGPT searches the web, it uses real time information, and the cutoff stops constraining the answer, because it comes from pages that exist today instead of from training data. Whether ChatGPT uses training data or real time information is decided per question. ChatGPT can search the web “to answer questions with current information”, and it “may search the web automatically when your question would benefit from current information”.

It “typically rewrites your query into one or more targeted queries” and may send more specific follow-ups after reviewing the first results. And the real time information you get back can still be wrong: OpenAI’s own caveat is that “search results and citations can be incomplete, outdated, or incorrect.”

The same capability sits across the other assistants: Gemini’s model pages list search grounding as supported, Claude’s model docs list web search, and Perplexity always searches. When search fires, the model’s memory is mostly out of the picture.

You can measure this. In our Credibility Layer study, 200 reasoning sessions ran 5,933 searches and produced 1,982 citations, and every recommendation came from a roughly 30-search credibility audit. So a buyer’s question that triggers search is answered from pages that exist right now, and what decides whether you’re in that answer is whether your pages are the ones the searches find.

What the cutoff means for your brand

The knowledge cutoff decides whether AI knows your brand: anything newer than the model’s cutoff doesn’t exist in its training data. A product you launched, renamed or repriced after the cutoff isn’t in the model’s memory, and if your brand is younger than the cutoff, training data can’t know you at all.

So the split worth checking, per buyer question, is whether the assistant searched. If it did, your live pages and the third-party pages it reads decide the answer. If it didn’t, the answer comes from memory, and only pre-cutoff reputation is in the running.

On your own pages that means current dates on the facts buyers ask about, pricing, versions, the team, and passages that answer the question outright, so a search that lands on your page can lift the answer. Off your site it means third-party sources the models already read carrying your claims, because a search-built answer is assembled from what the searches return.

What goes wrong in practice is predictable: assuming one date covers the whole product, quoting a cutoff without its source date, and treating “the model doesn’t know us” as permanent when it’s retrieval work.

Omnipresence records exactly this split. It tracks whether ChatGPT, Claude, Gemini and Perplexity recommend you, and for each answer it records whether the assistant searched, which pages it read and which it cited, so you can see which answers came from memory and which from retrieval, and which queries you’re losing either way. Its limits are real: it tracks four assistants, you bring your own model key, and its top-choice data is partial.

FAQ

Is ChatGPT’s knowledge up to date in real time?

No. Each model’s training data is fixed at its cutoff, and an answer is current only when the assistant searches, which it decides for itself, except Perplexity, which always searches. The same question can get a search-backed answer and a memory-backed answer on the same day.

What was GPT-4’s knowledge cutoff?

There was never one GPT-4 knowledge cutoff. GPT-4o and GPT-4o mini stop at October 1, 2023, and GPT-4 Turbo at December 1, 2023, on OpenAI’s model pages, and both lines are deprecated. Anyone quoting one GPT-4 cutoff is quoting one of several.

Does updating ChatGPT reset the cutoff?

No. Product updates don’t retrain the model, so the cutoff moves only when a new model ships, and a new model reaches some plans before others.

The dates will keep moving, because they belong to models, and models keep shipping under the same product names. Whether an assistant knows your brand is mostly a question of whether it searched before answering, and that you can check.