AI Model Knowledge Cutoff Dates: Every Published Figure, and Why They Disagree

A model's knowledge cutoff is the date its training data stops. Ask it about anything after that and it either says it does not know, retrieves the answer from live search, or — the expensive case — confidently tells you what was true before the cutoff.

Most "cutoff dates" lists you will find are a single column of months copied from somewhere else. That column is the problem. Across the five vendors below, the same model routinely carries two different published cutoffs, and which one is right depends on what you are asking: the broad span of data it was trained on, the narrower window its knowledge is actually reliable in, or what the product in front of you is serving today. Every vendor on this page publishes more than one figure for at least one of its models — and for Grok 4.7, two xAI surfaces give dates a month apart.

So this page gives every figure we could source to the vendor that owns the model, says which vendor surface each one came from, and flags the models where the vendor publishes nothing. Verified October 2026 — these pages change with every model release, so treat the vendor links as the source of truth rather than this table.

Key numbers

• Anthropic publishes two cutoffs per model on purpose. Its docs define a reliable knowledge cutoff — "the date through which the model's knowledge is most extensive and reliable" — separately from the training data cutoff, "the broader range of data used" (Anthropic model overview).
• Google's documented cutoff for the whole Gemini 3 family is January 2025, stated in the Gemini 3 developer guide, repeated on Google Cloud's individual model pages, and written into Google's own recommended system prompt (Gemini 3 developer guide).
• xAI's two surfaces disagree about the same model. The developer docs give Grok 4.7 a May 2026 cutoff; xAI's own model card for it, revised 21 September 2026, gives a pretraining cutoff of June 2026 and notes supplemental training data generated as late as August 2026 (xAI docs; Grok 4.7 model card).
• OpenAI's flagship GPT-5.6 models are the newest documented at February 16, 2026 — published on the Terra, Luna and Cyber model pages (OpenAI model docs).
• The oldest model still in wide use is years behind. GPT-4o's API page carries an October 1, 2023 cutoff, which is roughly three years of missing context on anything that changed since (OpenAI).
• Reported cutoffs are not the same as effective ones. A 2024 study of temporal bias in CommonCrawl-derived corpora found that models' effective knowledge cutoffs often differ from the dates their builders report, because of how the underlying crawls and deduplication work (Dated Data: Tracing Knowledge Cutoffs in Large Language Models).

The three things people call "the cutoff"

Before the tables, the distinction that makes them readable. Three different dates get reported as the cutoff, and vendors are not consistent about which one they publish:

1. Training data cutoff — the end of the span of data the model was trained on. The broadest figure, and usually the latest-sounding one.
2. Reliable knowledge cutoff — the date through which the model's knowledge is dense enough to trust. Always equal to or earlier than the training cutoff, because the last few months of any crawl are thin: the internet had not finished writing about them yet.
3. What the product serves today — the model actually answering you, which depends on your tier and on routing, and which can change without the documentation changing.

Anthropic is the only vendor that publishes (1) and (2) side by side as named columns. Google publishes a family-level figure and then hedges it per domain on individual model cards. xAI publishes one figure in its docs and a different, more detailed one on its model cards. OpenAI publishes a single date per model page and does not label which kind it is.

This is why a one-column table is wrong by construction, and why every row below names its source.

OpenAI

From OpenAI's own model pages. These are published as a single "knowledge cutoff" per model, without a training-versus-reliable distinction.

Source: OpenAI model documentation.

Three notes on precision, because this is the vendor where guessing is most tempting.

GPT-5.6 Sol has no published cutoff that we could find. Its model page gives the context window and max output tokens but no cutoff. Sol is the flagship, so it is the single figure most readers want, and the obvious move is to assume it shares the February 16, 2026 date its three siblings carry. We are not doing that: OpenAI's comparison view also surfaces April 30, 2026 and May 18, 2026 against models it does not clearly name, and either could belong to Sol. An unlabelled date is not a source.

GPT-4o is the clearest case of a model with two true answers. The API model page says October 1, 2023. But OpenAI's January 2025 ChatGPT update extended GPT-4o's training data to June 2024 for the ChatGPT product. Same model name, two different cutoffs, depending on whether you meet it through the API or the consumer app. GPT-4o, GPT-4.1, GPT-4.1 mini and o4-mini were then retired from ChatGPT on February 13, 2026 (model release notes).

The ChatGPT you are using is probably not the model you picked. Which model serves a given request depends on tier and routing, so the cutoff that applies to your answer is not always knowable from the outside. That is one of several reasons ChatGPT gives different answers to the same question.

Anthropic

Anthropic publishes both figures as named columns, and defines them explicitly: the reliable knowledge cutoff is "the date through which the model's knowledge is most extensive and reliable", while the training data cutoff is "the broader range of data used".

Sources: the model overview table for the current lineup and both columns; Anthropic's support article on training data recency for the full list including retired models; and the Transparency Hub for per-model detail.

Worth knowing: on the current lineup both columns read June 2026, so the two-figure distinction currently makes no practical difference for Anthropic's newest models. It mattered more on earlier generations, where the reliable window closed months before the training span did. The three Anthropic surfaces agree with each other wherever they overlap — the Transparency Hub simply has not added rows for the 5.5 generation yet, which is a gap in coverage rather than a conflict.

Google

Google's documentation treats the cutoff as a property of the model family, not of each model.

Sources: Gemini 3 developer guide, Google Cloud's per-model documentation, and the DeepMind model cards.

The two Flash cards look at first like Google contradicting Google, and that is how it is often reported. Read properly, they do not: the cards give March 2026 as the cutoff for some domains while stating that in other domains the model's knowledge is limited to January 2025, in line with the Gemini 3 model family. That is a deliberate hedge about uneven coverage, not a competing claim — but it does mean there is no single correct cell for those two models, which is exactly the shape of problem this page exists to show.

Google's own guidance for handling it is instructive: its prompting documentation recommends telling the model its own cutoff in the system prompt — literally "Your knowledge cutoff date is January 2025" — and routing anything newer through Search Grounding rather than relying on training knowledge at all.

Meta (Llama)

Meta publishes a pretraining data cutoff on each Llama model card. Note the label: this is the training-span figure, with no reliable-knowledge equivalent.

Sources: the Llama 4 model card and the meta-llama model cards on Hugging Face, Meta's official distribution channel — for example Llama 3.3 70B Instruct.

Llama 2 is the one row that shows the pattern plainly: pretraining stops in September 2022, but some tuning data runs to July 2023. One model, two dates, both published by the vendor.

xAI (Grok)

Sources: xAI models documentation and the Grok 4.7 model card.

This is the genuine contradiction in the set, and we are not resolving it for you. Two xAI surfaces give different cutoffs for the same model, a month apart, and the model card is both the more detailed and the more recently revised of the two. If you need a number for Grok 4.7, cite the surface you took it from.

What we could not source

Being explicit about the gaps matters more than a longer table, because a row nobody can source is a row that ends up quoted back as fact.

• Mistral does not appear to publish a per-model knowledge cutoff in its documentation. Its API docs mention a cutoff only as something you might include in your own system prompt.
• DeepSeek — we found no cutoff date in its published model documentation.
• GPT-5.6 Sol, as above: OpenAI publishes the model page but not a cutoff on it.

Absence of a published cutoff is not evidence that a model is current. It means you cannot plan around the date, and should test the model's knowledge of recent events directly instead.

Reported cutoff versus effective cutoff

Even a correctly quoted date is an approximation of what the model actually knows, for a reason that has nothing to do with vendor honesty.

Training corpora are assembled from web crawls, and crawls are temporally lumpy. A page published in the last month before a cutoff has had almost no time to be linked, quoted, mirrored or discussed, so it appears in the corpus once or twice rather than hundreds of times. Events from eighteen months before the cutoff are saturated. The result is a long fade rather than a clean edge, and the 2024 Dated Data analysis found that effective cutoffs frequently diverge from reported ones as a consequence of exactly this, plus deduplication choices.

This is precisely what Anthropic's two-column approach is describing, and it is the practical reason the reliable figure is the one to plan against where a vendor publishes one. For everyone else, assume the real edge sits somewhat earlier than the published date.

Why this matters for whether AI mentions your brand

The cutoff decides which of two completely different mechanisms has to carry your brand.

Inside the cutoff, a model may know you from training. It can name you with no retrieval at all, which is the most durable kind of AI visibility there is — it does not depend on a crawl, an index or a citation.

Outside the cutoff, training knowledge cannot help you, and everything depends on live retrieval. Google's answer is Search Grounding; OpenAI's is ChatGPT search, which runs through a different crawler from the training one. Which means that for anything newer than the dates above — a company founded last year, a rebrand, a repositioning, a new pricing model, a product line that did not exist — you are invisible in the training layer by definition and entirely dependent on being retrievable.

Two practical consequences:

• If your category story changed recently, the models are still telling the old one. A model with a January 2025 cutoff describes your 2025 positioning, your old pricing, and competitors you have since overtaken. No amount of on-site optimisation changes what is already frozen in the weights.
• Retrieval is the only lever you control on a useful timescale. Which is why the crawler-level work matters: being eligible for the retrieval layer is what AI crawler user agents is about, and where ChatGPT gets its data covers how the three layers interact.

Where Sightivo fits

Knowing the cutoff tells you which mechanism you are fighting. It does not tell you whether it is working — whether an assistant actually names you today, for the questions your buyers ask.

Sightivo runs your buyer prompts against ChatGPT and Claude, records which brands each answer names and which sources it drew on, and turns the gap into a ranked list of pages to get mentioned on. That is the measurement that tells you whether the training layer, the retrieval layer, or neither is carrying you right now. The free AI visibility checker shows in a minute whether ChatGPT names you for five prompts in your category.

Common Questions

What is ChatGPT's knowledge cutoff?

It depends on the model. OpenAI's model pages give February 16, 2026 for GPT-5.6 Terra, Luna and Cyber; December 1, 2025 for GPT-5.5; August 31, 2025 for GPT-5.4; September 30, 2024 for GPT-5 and GPT-5.1; and October 1, 2023 for GPT-4o on the API. GPT-5.6 Sol, the flagship, has no cutoff published on its model page. Because ChatGPT routes requests between models by tier and load, the cutoff that applied to any given answer is not always knowable from the outside.

What is Claude's knowledge cutoff?

June 2026 for the current lineup — Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 — for both the reliable knowledge cutoff and the training data cutoff. Earlier models run back from there: May 2026 for Opus 5, January 2026 for Sonnet 5, Fable 5, Opus 4.8 and Opus 4.7, August 2025 for Sonnet 4.6 and Opus 4.6, July 2025 for Haiku 4.5, and August 2023 for Opus 3.

What is Gemini's knowledge cutoff?

Google documents January 2025 for the Gemini 3 family as a whole, and repeats it on individual Google Cloud model pages and in its recommended system prompt. The DeepMind model cards for Gemini 3.6 Flash and 3.7 Flash give March 2026 for some domains while stating that in others knowledge is limited to January 2025, in line with the family. So there is no single correct figure for those two models.

Why do different sources give different knowledge cutoff dates for the same model?

Usually because they are quoting different things. Three distinct dates get called "the cutoff": the training data cutoff, the narrower reliable knowledge cutoff, and what the product is actually serving today. Anthropic publishes the first two as separate columns. Google publishes a family figure then hedges it per domain. xAI's developer docs and its own model card give different dates for Grok 4.7. And GPT-4o carries October 1, 2023 on the API while OpenAI extended the ChatGPT version's data to June 2024. Always note which surface a date came from.

What is the difference between a training data cutoff and a knowledge cutoff?

The training data cutoff is the end of the span of data used to train the model. The reliable knowledge cutoff, in Anthropic's definition, is the date through which the model's knowledge is "most extensive and reliable" — it is equal to or earlier than the training cutoff, because the final months of any crawl are thinly represented. If a vendor publishes both, plan against the reliable figure.

Does a model's knowledge cutoff mean it knows nothing after that date?

Not in practice, for two reasons. Coverage fades rather than stopping dead, so a model often has patchy knowledge of events shortly after its stated date, and research on CommonCrawl-derived corpora has found effective cutoffs that differ from reported ones. And most assistants can retrieve live information: Google routes newer questions through Search Grounding and ChatGPT through its own search, so the answer you get may come from today's web rather than from training at all.

How does the knowledge cutoff affect whether AI tools mention my brand?

It decides which mechanism has to carry you. Inside the cutoff, a model may name you from training alone, with no retrieval needed — the most durable form of AI visibility. Outside it, training cannot help and you depend entirely on being retrievable through live search. So anything about your company that changed after the cutoff — a rebrand, new pricing, a new product line — reaches an assistant only through the retrieval layer.

How often do these dates change?

Every time a vendor ships a model, which in 2026 has meant every few weeks. The dates for existing models change too: OpenAI extended GPT-4o's ChatGPT cutoff after release, and xAI revised its Grok 4.7 model card in September 2026. Check the vendor page rather than a third-party list, including this one, before you rely on a specific date.

Key Takeaways

• There is no single "cutoff" per model. Three different dates get the name — training span, reliable knowledge, and what the product serves today — and most vendors publish only one without saying which.
• Anthropic is the clearest: it publishes a reliable knowledge cutoff and a training data cutoff side by side, both June 2026 across the current lineup.
• Google treats the cutoff as a family property — January 2025 for Gemini 3 — and hedges it per domain on two Flash model cards, which is why those models have no single correct figure.
• xAI genuinely contradicts itself: May 2026 in the developer docs versus June 2026 pretraining on the model card for Grok 4.7.
• OpenAI's newest documented figure is February 16, 2026 for three GPT-5.6 models, and the flagship Sol publishes no cutoff at all.
• The reported date overstates what the model reliably knows, because the last months before any cutoff are thinly represented in the training corpus.
• For AEO, the cutoff tells you whether training knowledge or live retrieval has to carry your brand. Anything that changed recently depends entirely on retrieval.

Start with the free AI visibility checker to see whether the assistants name you today. Then read where ChatGPT gets its data for how training, licensing and live search combine into one answer, make sure you are eligible for the retrieval layer with AI crawler user agents, and compare monitoring options in the best AI visibility tools.