AI Knowledge Cutoff Dates: ChatGPT, Claude, Gemini and More (Updated October 2026)

Every few weeks someone asks an AI chatbot about a product that launched last month and gets back either a confident wrong answer or an honest "I don't have information on that." Which one happens depends on a specific, documented fact about the model answering: its knowledge cutoff, and whether the product in front of you gave it a way to look past that cutoff before answering.

This post is a reference table, pulled from the vendors' own current model documentation, plus a plain explanation of what a cutoff actually means and why two AI products can give wildly different answers about the same recent event. Last checked October 13, 2026. Model lineups change often enough that you should treat this as a snapshot, not a permanent reference, and the source links are there so you can re-check it yourself.

What a knowledge cutoff actually means

A large language model learns everything it knows during training, on a fixed dataset assembled up to some point in time. After that point, training stops, and the model is frozen. It doesn't learn anything new unless a vendor runs another training cycle and ships a new model version. The date marking the edge of that training data is the knowledge cutoff, sometimes called a training data cutoff.

That's different from whether the product you're using can access anything past that date. Most consumer AI products today layer a retrieval or search capability on top of the base model: ChatGPT can browse the web, Gemini can ground its answers in Google Search results, Claude has a web search tool, and Grok has a tight integration with live X posts. When one of those products answers a question about something from last week, it's not because the underlying model's cutoff moved. It's because the product fetched fresh information at the moment of your question and handed it to the model as context, which the model then summarizes or reasons over. Turn off web search or grounding, and the same model reverts to answering from training data alone, cutoff and all.

This is the single most common confusion in how people talk about this. "ChatGPT knows about the news" and "GPT's training data includes the news" are two different claims, and only one of them is usually true at a given moment.

Current flagship models and their documented cutoffs

The table below reflects what each vendor states on its own current model documentation. Where a vendor hasn't published a cutoff, that's stated plainly rather than guessed at.

A few things in that table are worth calling out directly rather than letting them pass as a footnote.

Anthropic is the only vendor here that publishes two separate dates for the same model: a "training data cutoff" and a "reliable knowledge cutoff." Its own documentation defines the reliable knowledge cutoff as "the date through which the model's knowledge is most extensive and reliable," while the training data cutoff is described as the broader range of data used in training. For the current Claude lineup, both happen to land on the same month, but Anthropic's own phrasing makes clear these aren't guaranteed to be identical in every release, and the company points readers to its Transparency Hub for the full detail behind each number.

Google is the clearest gap in this table. Its Gemini API model documentation lists current models (Gemini 3.1 Pro and Gemini 3 Flash were both marked Preview at the time this was checked) without a knowledge cutoff field anywhere on the page. That's a real absence, not an oversight on our end; if Google publishes this elsewhere, it isn't on the model listing page developers would normally check.

Mistral's documentation uses version-dated model names (Mistral Large 4 is tagged "v26.10," for instance) that look like they might be cutoff dates at a glance. They're not. A version tag tells you when that model build was released, not how recent its training data is, and Mistral's current docs don't separately state the latter.

xAI publishes a cutoff for its newest model, Grok 4.7, but the same documentation doesn't carry that field forward consistently for the earlier versions still listed as current (4.6, 4.5, 4.3). That's an easy trap: seeing one model in a vendor's lineup with a clearly stated cutoff doesn't mean every sibling model in that same table has one too.

Why answers about recent events differ by product

Put the cutoff table next to the grounding question and the differences between products stop looking random. A product built on a model with a June 2026 training cutoff, with no live web search enabled, is working from the same frozen snapshot of the world no matter how recent the actual conversation is. A product built on an older model but with aggressive, always-on web grounding (Gemini's Search grounding and Perplexity's architecture both work this way by design) can describe something that happened an hour ago, because it's not relying on memorized training data for that answer at all.

This is also why the same underlying model family can feel "smarter about current events" in one surface than another. ChatGPT with web browsing turned on and ChatGPT answering from memory alone are, in a meaningful sense, two different research tools wearing the same name. The cutoff date tells you the ceiling of the second one. It tells you almost nothing about the first one, because the first one isn't really relying on that ceiling for the answer you just got.

If you want to understand the retrieval side of this more precisely, query fan-out explained covers how Google's AI Mode decomposes a question into multiple live searches rather than answering from a frozen model alone, which is the same underlying mechanism making grounded answers current regardless of the base model's cutoff.

What this means for your brand

A knowledge cutoff isn't trivia. It has a direct, practical consequence for anyone doing AI-visibility or AEO work: if your product, rebrand, pricing change, or company name launched after a model's training cutoff, that model has no idea it exists, full stop, when it's answering from memory alone. Ask a non-grounded model about your brand and you might get silence, a confident guess based on a similarly named competitor, or an answer describing your product as it existed in an earlier, outdated state. None of those are the model being unreliable in some general sense. They're the model doing exactly what a frozen snapshot should do: describing the world as it looked on the day training stopped.

The fix isn't waiting for the next training run, which you don't control and can't schedule around. It's making sure the grounded, retrieval-based layer most products now sit on top of their base model can find and cite accurate, current information about you at query time, regardless of what the underlying model memorized. That means the same groundwork that gets you cited in any AI answer: clean structured data, a presence on sources these systems actually retrieve from, and content that directly, specifically answers the questions people are asking about you right now. How to get cited by ChatGPT and Google AI Overviews and LLM seeding both cover that groundwork in more depth, and how to get cited by Claude covers the same problem specifically for a product whose web search tool is a separate, optional layer rather than always-on by default.

The practical checklist, if a new product or a meaningful change just shipped: assume every model's base knowledge is wrong or missing about it until proven otherwise, verify whether the product surface you care about (ChatGPT, Gemini, Claude, Perplexity, Copilot) is actually grounding that specific type of query with live search before you worry about training data at all, and prioritize getting accurate information into the sources those grounded answers pull from today rather than waiting on a future training cutoff to catch up on its own.

FAQ

Does a newer knowledge cutoff mean a model is automatically better?

No. Cutoff date measures recency of training data, not reasoning quality, accuracy, or usefulness for a given task. A model with an older cutoff but strong grounding and web search can answer current-events questions more reliably in practice than a model with a newer cutoff and no retrieval layer at all, because the grounded one isn't relying on memorized training data for that specific answer.

Why don't Google, Mistral, and DeepSeek publish a cutoff date the way OpenAI and Anthropic do?

Vendors aren't required to publish this, and the reasons for not doing so vary (frequent incremental updates to training data, competitive reasons, or simply not treating it as a headline spec the way context window or pricing are treated). The honest answer, when a vendor's current documentation doesn't state one, is "not published," not a guess based on release date or version number.

If a model doesn't have a published cutoff, does that mean it has no cutoff at all?

No. Every model trained on a fixed dataset has some point at which that dataset stops, even if the vendor hasn't disclosed the exact date publicly. "Not published" means the information isn't available in the vendor's own current documentation, not that the model somehow continuously learns in real time (outside of a retrieval or grounding layer bolted on top, which is a separate thing from the base model's training).

How often should I re-check this table myself instead of trusting a cached version?

Model lineups from every major vendor here have shipped new flagship versions multiple times within a single year recently, and a vendor can also quietly update a documentation page without a major model release. Check the vendor's own current model page directly before relying on a specific cutoff date for anything client-facing, and treat any third-party table, including this one, as a starting point rather than a permanent source.

Bookmark the vendor pages linked in the table above rather than this post's numbers specifically. The exact dates here will be outdated within months; the method for checking them, and the distinction between a frozen training cutoff and a live grounding layer, won't be.