knowledge cutoff
The last date covered by a language model’s training data, beyond which the model knows little or nothing unless newer information is supplied to it.
- ChatGPT’s September 2021 cutoff, stated in OpenAI’s 2023 example prompt
- GPT-4 Turbo’s April 2023 cutoff, announced at OpenAI’s DevDay
- Gemini 2.5 Pro’s January 2025 cutoff, ahead of a June 2025 update
- Claude Opus 5’s May 2026 cutoff, for a July 2026 release
- Current meaning
- 2024
- By 2024, model releases commonly stated a knowledge cutoff date.
- Also called
- Training cutoff
- Anthropic’s 2025 prompt for Claude Sonnet 3.7 says training cutoff date.
In plain words
In AI, a knowledge cutoff is the date a model’s training data ends; after it, the model knows little unless newer facts are supplied.
A large language model learns from a huge collection of text gathered up to a certain date. That date is its knowledge cutoff. The model reaches the public later, often months later. Google lists Gemini 2.5 Pro’s knowledge cutoff as January 2025 and its latest update as June 2025.1
Once training ends, what the model knows stays fixed. It can speak about later events only if someone supplies the facts, for example through a web search.2 Otherwise it answers from old information. A price, a law or an officeholder may have changed since, so an answer that was once right is now wrong.
A model also has no clock. Anthropic’s instructions for its Claude chatbot state the current date and name the end of the model’s data as its training cutoff date.3 Without such help a model can lose track of time. In November 2025, Google’s Gemini 3 insisted to Andrej Karpathy, an AI researcher, that the year was 2024. It changed its mind only when he switched on its search tool.
The stated date is approximate. Anthropic’s prompt for Claude Sonnet 4 set a reliable knowledge cutoff at the end of January 2025. After that point, its answers stop being reliable.4 Anthropic’s model table, however, listed training data through March 2025. A 2024 study found that the knowledge models actually show often ends at a different date from the one their makers report.5
An outdated answer differs from a hallucination. The outdated fact was true when the data was gathered; a hallucinated one never was.
How it works
Builders first gather training text, such as web pages and books, up to a chosen date. The model is then pretrained on that text. Training adjusts its weights, the numbers that store what it learns, until it predicts each next word well. Once training ends the weights are fixed, and answering questions does not change them.
The cutoff is not a clean line. Cheng and colleagues traced two reasons that the knowledge a model shows can end at a different date from the reported one. Newer snapshots of the web, such as Common Crawl dumps, still hold a good deal of older material. Filters meant to remove duplicate documents also struggle with near-copies.5
The model has no way to know the date unless it is told. The system prompt, the instructions a product sends with each conversation, can state it. OpenAI’s 2023 example prompt for ChatGPT gave both the knowledge cutoff and the current date.
Newer facts have to arrive the same way, through the input. A search tool or retrieval-augmented generation fetches current text and places it in the model’s context window. IBM describes this as giving a model the most current facts without retraining it.6
- GPT-4 Turbo, OpenAI
- Cutoff April 2023, announced November 2023
- Gemini 2.5 Pro, Google
- Cutoff January 2025, latest update June 2025
- Claude Sonnet 4.6, Anthropic
- Cutoff May 2025, released February 2026
- Claude Opus 5, Anthropic
- Cutoff May 2026, released July 2026
Origin and first known use
Cutoff is a noun formed from the verb cut off, and it is also spelled cut-off. Merriam-Webster dates it to 1647. In AI the phrase names the date on which a model’s training data ends. OpenAI’s 2023 guide to the ChatGPT API showed the model being told its knowledge cutoff in its instructions. By 2024 researchers noted that model releases commonly stated a cutoff date. Anthropic also uses reliable knowledge cutoff for the date after which a model stops answering reliably. Both spellings appear in AI use, at times in a single sentence.
- 1647 Merriam-Webster dates the noun cutoff, also spelled cut-off, to 1647, when it meant the act of cutting off.Merriam-Webster, “cutoff” merriam-webster.com →: source for 1647
- 2021 Lazaridou and colleagues find that language models predict text from after their training period increasingly badly over time, and that larger models alone do not fix this.Lazaridou et al., “Mind the Gap: Assessing Temporal Generalization in Neural Language Models”, NeurIPS 2021 arXiv 2102.01951 →: source for 2021
- 2021 Komeili, Shuster and Weston note that a model’s facts are frozen at the point of training, and train a chatbot to search the internet instead.Komeili, Shuster and Weston, “Internet-Augmented Dialogue Generation”, ACL 2022 arXiv 2107.07566 →: source for 2021
- 2023 OpenAI’s ChatML guide for the new ChatGPT API shows a system message that gives the model its knowledge cutoff, September 2021, and the current date.OpenAI, “Chat Markup Language (ChatML for short)”, openai-python repository, 1 March 2023 github.com →: source for 2023
- 2023 OpenAI announces GPT-4 Turbo and says it has knowledge of world events up to April 2023.OpenAI, “New models and developer products announced at DevDay”, 6 November 2023 openai.com →: source for 2023
- 2024 Cheng and colleagues note that model releases commonly state a knowledge cutoff date, and find that the knowledge models show often ends at a different date.Cheng et al., “Dated Data: Tracing Knowledge Cutoffs in Large Language Models” arXiv 2403.12958 →: source for 2024
- 2025 Anthropic’s system prompt for Claude Sonnet 4 sets a reliable knowledge cutoff, the end of January 2025, after which its answers stop being reliable.Anthropic, system prompt for Claude Sonnet 4, 22 May 2025 platform.claude.com →: source for 2025
- 2025 Google’s Gemini 3, trained on data only through 2024, insists to Andrej Karpathy that the year is 2024 until he turns on its Google Search tool.Julie Bort, TechCrunch report on Gemini 3 and Andrej Karpathy, 20 November 2025 techcrunch.com →: source for 2025
In the wild
-
“You are ChatGPT, a large language model trained by OpenAI. Answer as concisely as possible. Knowledge cutoff: 2021-09-01 Current date: 2023-03-01”
OpenAI, example system message in its ChatML guide github.com →: OpenAI, example system message in its ChatML guide, 1 Mar 2023Archived →: copy of OpenAI, example system message in its ChatML guide, 1 Mar 2023 -
“Released Large Language Models (LLMs) are often paired with a claimed knowledge cutoff date, or the dates at which training data was gathered.”
-
“…Claude tells the user the most recent information per its knowledge cutoff and informs them things may have changed since the knowledge cut-off.”
Anthropic, system prompt for Claude Sonnet 4 platform.claude.com →: Anthropic, system prompt for Claude Sonnet 4, 22 May 2025Archived →: copy of Anthropic, system prompt for Claude Sonnet 4, 22 May 2025 -
“Anthropic's model comparison table lists a training data cut-off of March 2025 for both Opus 4 and Sonnet 4…”
-
“Gemini 3 refused to believe it was 2025, and hilarity ensued”
-
“Claude Opus 5 has a knowledge cutoff date of May 2026.”
Anthropic, Transparency Hub model report anthropic.com →: Anthropic, Transparency Hub model report, 17 Aug 2026Archived →: copy of Anthropic, Transparency Hub model report, 17 Aug 2026
Not to be confused with
-
Context window
The context window holds what is sent with the current request, and nothing in it carries over. The knowledge cutoff dates what the model learned in training, which stays fixed until a new version is trained.
-
Hallucination
A hallucinated fact was never true. An answer limited by the knowledge cutoff was true when the training data was gathered and has changed since. Asked about later events, a model may also guess and hallucinate.
-
Release date
The day a model becomes publicly available. It always falls after the knowledge cutoff, often by months: Anthropic released Claude Sonnet 4.6 in February 2026 with a cutoff of May 2025.
Questions
What is a knowledge cutoff in AI?
It is the last date covered by the data a language model was trained on. The model knows little or nothing about later events unless they are supplied to it, for example by a web search. Model makers usually state the date. Google lists January 2025 for Gemini 2.5 Pro, and Anthropic lists May 2026 for Claude Opus 5.
Why do AI chatbots not know about recent events?
A language model learns from text gathered up to its knowledge cutoff, and what it knows stays fixed once training ends. Models also tend to reach the public months after their cutoff: Anthropic released Claude Sonnet 4.6 in February 2026 with a cutoff of May 2025. Newer facts reach a model only through its input, such as web search results.
Why does an AI chatbot think it is a different year?
A model has no clock, and its training text ends at the knowledge cutoff. Unless it is told the date, it cannot tell how much time has passed. OpenAI’s 2023 example instructions for ChatGPT and Anthropic’s prompts for Claude both supply the current date. In November 2025, Google’s Gemini 3 insisted to Andrej Karpathy that the year was 2024, until he turned on its search tool.
Is a model’s stated knowledge cutoff accurate?
Only roughly. A 2024 study by Cheng and colleagues found that the knowledge models actually show often ends at a different date from the reported one. They traced this to older material in newer web crawls and to problems removing duplicate documents. Anthropic gave Claude Sonnet 4 a reliable knowledge cutoff of January 2025, while its training data ran to March 2025.
How can a model answer questions about events after its knowledge cutoff?
Only if the facts are supplied when the question is asked. A search tool or retrieval-augmented generation finds current text and places it in the model’s context window, and the model answers from it. IBM describes this as giving a model the most current facts without retraining it. Without such help, the model answers from training data that may be out of date.
Watch
Cite this entry
“Knowledge cutoff.” SUPERR_dictionary, Superrduperr, 24 Sept. 2026, superrduperr.com/
Superrduperr. (2026, September 24). Knowledge cutoff. In SUPERR_dictionary. https://superrduperr.com/
Superrduperr. “Knowledge cutoff.” SUPERR_dictionary. September 24, 2026. https://superrduperr.com/