# knowledge cutoff

Also called: training cutoff, training data cutoff, reliable knowledge cutoff, data cutoff, cutoff date

A knowledge cutoff is the date after which a model's training data stops, so the model has no record of later events, releases or API changes unless you put them in the prompt. Providers publish it per model, and some publish two dates: Anthropic lists a reliable knowledge cutoff, through which coverage is most extensive, and a training data cutoff for the full range of data used. The model has no clock, so it cannot tell how far today is from either date.

## What does a knowledge cutoff mean in practice?

A model is trained on text collected up to some date, then its weights are frozen and it is served for months or years. The knowledge cutoff is that collection date. Every fact the model can recall without help comes from text written before it, so a library released last month, a renamed API parameter or this week's news are absent from the weights.

Providers print the date on each model's spec page. OpenAI's model pages carry a single Knowledge cutoff field. Anthropic's models overview has two rows: a reliable knowledge cutoff, defined as the date through which the model's knowledge is most extensive and reliable, and a training data cutoff, the broader range of data used in training. When the two differ, the gap between them is a stretch the model has seen some text about, but less of it.

Marketing copy tends to read the date as "the model knows everything up to here". In code you should read it as an upper bound: past that date the model knows nothing, and the months before it may be thinly covered.

## Why the reported cutoff and the effective cutoff differ

Cheng et al. (2024), in Dated Data: Tracing Knowledge Cutoffs in Large Language Models, measured the cutoff a model behaves as if it has, per resource and per topic, and found that these effective cutoffs often differ from the date the provider reports. They trace the gap to two causes in how pretraining corpora are built: new CommonCrawl dumps still contain older pages, and lexical and semantic deduplication complicate which copies of a document survive into training.

The published date describes the dataset as a whole, not any one subject. If your product depends on a specific framework, standard or price list, test what the model knows about that subject directly instead of trusting the cutoff on the spec page.

## What changes in your code?

Pass the date. The Messages and Chat Completions APIs do not tell the model what day it is. Put today's date in the system prompt whenever the answer depends on it: deadlines, ages, "latest version" questions, relative dates like "next Tuesday".Pass the facts. For anything newer than the cutoff, retrieve current text and put it in the context window with RAG, or give the model a search tool. Anthropic documents its web search tool as the way to answer with information beyond the knowledge cutoff.Expect stale code. Generated code targets the SDK versions and parameters the model saw in training. Pin dependency versions, and give the model the current docs for any API that changed after the cutoff.Put post-cutoff questions in your evals. A few rows about recent changes show whether the model searches or says it is unsure, rather than answering confidently from old data.

## Passing today's date and a search tool to a model with a fixed cutoff

import datetime, anthropic

client = anthropic.Anthropic()
today = datetime.date.today().isoformat()   # "2026-10-12"

resp = client.messages.create(
    model="claude-opus-5-5",   # reliable knowledge cutoff: Jun 2026
    max_tokens=1024,
    system=(
        f"Today's date is {today}. Your training data ends before this date. "
        "Search before answering anything that may have changed since."
    ),
    messages=[{"role": "user",
               "content": "What is the latest stable release of Python?"}],
    tools=[{"type": "web_search_20250305", "name": "web_search", "max_uses": 3}],
)

for block in resp.content:
    print(block.type)

# Block order when the model decides to search:
# text
# server_tool_use          <- the query the model wrote
# web_search_tool_result   <- results, each with a page_age
# text                     <- the answer, carrying citations

The system prompt gives the model the date it cannot know, and the web search tool gives it a path to facts newer than its cutoff. The model decides whether to search; when it does, the response carries a server_tool_use block and a web_search_tool_result block before the cited answer, as Anthropic's web search docs describe.

## Frequently asked

Q: Why does the model think it is still the year of its training data?
A: The model has no clock and the API does not send it the date. Without a date in the prompt, it falls back on the most recent dates in its training data. Put today's date in the system prompt and the problem goes away for date arithmetic, though the model still lacks events after its cutoff.

Q: Can I ask the model what its knowledge cutoff is?
A: You can, and you should not rely on the answer. The reply is generated text like any other, assembled from whatever the training data said about the model, and Cheng et al. found that even provider-reported dates differ from effective ones. Read the cutoff from the provider's model page and test your specific topics.

Q: Which date should I use when a provider lists two cutoffs?
A: Plan around the reliable knowledge cutoff. Anthropic defines it as the date through which the model's knowledge is most extensive and reliable, while the training data cutoff only marks the outer edge of the data used. Treat anything between the two as partially known and supply current sources for it.

Q: Does RAG or a search tool move the knowledge cutoff?
A: No. The weights stay the same and so does the cutoff. Retrieval and search put current text into the context window for one request, and the model reads it there. The next request without that text is back to what the weights know.

## Sources

- Anthropic docs: Models overview: https://platform.claude.com/docs/en/models/overview
  per-model reliable knowledge cutoff and training data cutoff rows, with the footnote defining the difference between them
- OpenAI docs: Models: https://developers.openai.com/api/docs/models
  model catalogue whose per-model pages list a single Knowledge cutoff field
- Dated Data: Tracing Knowledge Cutoffs in Large Language Models (Cheng et al., 2024): https://arxiv.org/abs/2403.12958
  shows effective cutoffs often differ from reported ones and vary by resource, attributing the gap to old data in new CommonCrawl dumps and to deduplication
- Anthropic docs: Web search tool: https://platform.claude.com/docs/en/agents-and-tools/tool-use/web-search-tool
  the server tool for answering with information beyond the knowledge cutoff, its tool definition and its response block structure

## Related terms

- model: https://ai.developertools.io/model/index.md
- hallucination: https://ai.developertools.io/hallucination/index.md
- RAG: https://ai.developertools.io/rag/index.md
- tool calling: https://ai.developertools.io/tool-calling/index.md
- prompt: https://ai.developertools.io/prompt/index.md
- eval: https://ai.developertools.io/eval/index.md

Canonical page: https://ai.developertools.io/knowledge-cutoff/
Edited by Derek Wonk. Last updated 2026-10-12.