What is a training cutoff?

Where the cutoff comes from

A language model is built in stages. First, companies collect training data: mostly web pages (often starting from the Common Crawl archive5), plus books, code and licensed or purchased data. GPT-3’s paper, for example, describes filtered Common Crawl data from 2016 to 20194. At some point the collection stops. That moment is the training cutoff.

Then the model is trained on that frozen snapshot, which takes weeks to months, and afterwards tuned, safety-tested and packaged into products. So a model is always released months after its cutoff, and then used for a year or more. Anthropic, for example, lists retirement dates at least a year after release1. Nothing the model learned is updated in the meantime: its knowledge is fixed in its weights.

The gap for today’s models

From our model registry: cutoff, release date and the gap between them (then add however long you’ve been using it).

Cutoffs as published by each lab12. Full list by model: briefings by model.

Four assistants, one scale

What each assistant knows (graphite), the months between its cutoff and its launch (hatched), and the news since (pink). The numbers count the events on this site after each cutoff.

The cutoff is blurrier than one date

The internet keeps writing about an event for months and years after it happens: analysis, follow-ups, Wikipedia edits. So at training time the last few months before the cutoff are only thinly covered, and the model knows them less reliably than older events. Anthropic publishes two dates for this reason, a “training data cutoff” and an earlier “reliable knowledge cutoff”. For Claude Haiku 4.5 they are July 2025 and February 20251.

Researchers have also found that a model’s effective cutoff often differs from the reported one, partly because web crawls contain older copies of pages and because of how data is deduplicated7. And models get measurably worse at text from after their training period, more so the further it lies in the future6. Time even leaves a trace inside the model’s weights11.

In our own test, asked without tools, Claude Opus 5.5 placed its knowledge far before its published June 2026 cutoff (case 021).

Why models don’t know today’s date

A model has no clock. On its own it tends to assume that “now” is roughly its training cutoff. Chat apps work around this by quietly giving the model the current date in a hidden instruction (the system prompt) at the start of every conversation3. Through the raw API, or in tools that don’t do this, the model may not know what day it is at all.

Why models call real news fake

Knowing the date isn’t enough. When a model meets news from after its cutoff, such as a new pope, a new president, or a newer AI model with an unfamiliar name, it often concludes the news must be invented, rather than that time has passed. Real examples:

• 2023: Bing Chat insisted it was still 2022 and that a film already in cinemas wasn’t out yet. Case 002
• 2025: several chatbots said “Pope Leo XIV does not exist” months after his election. Case 011
• 2025: an early build of Gemini 3 refused to believe it was 2025 and called real news articles AI-generated. Case 012
• 2026: we gave Gemini 3.1 Pro our own sourced timeline. Offline it called it fiction in 9 of 10 runs, even when told the date; with Google Search on it said “I’m stunned!” and accepted it. Case 001d

We call this cutoff blindness. All 28 documented cases, with sources: Called fake.

How to fix it

1. Turn on web search when your chat app offers it. Letting a model look things up (retrieval) is the standard way to give it fresh facts9, and search results measurably improve answers about current events810. In our test, search was the only thing that reliably worked.
2. Tell it today’s date if it might not know (most apps already do).
3. Give it a briefing of what it missed. That’s what this site is for: paste postcutoff.com/ai into the chat, or pick the file for your model on briefings by model.

Sources

1. Anthropic, Models overview: “Reliable knowledge cutoff” and “Training data cutoff” per modeldocs.claude.com
2. OpenAI, Models: each model page lists its knowledge cutoffplatform.openai.com
3. Anthropic, System prompts: the apps give Claude “up-to-date information, such as the current date”docs.claude.com
4. Brown et al. (2020), Language Models are Few-Shot Learners (GPT-3; training data from Common Crawl 2016–2019)arxiv.org
5. Common Crawl, the open web archive most training sets start fromcommoncrawl.org
6. Lazaridou et al. (2021), Mind the Gap: Assessing Temporal Generalization in Neural Language Modelsarxiv.org
7. Cheng, Marone, Weller et al. (2024), Dated Data: Tracing Knowledge Cutoffs in Large Language Modelsarxiv.org
8. Kasai et al. (2022), RealTime QA: What’s the Answer Right Now?arxiv.org
9. Lewis et al. (2020), Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasksarxiv.org
10. Vu et al. (2023), FreshLLMs: Refreshing Large Language Models with Search Engine Augmentationarxiv.org
11. Nylund, Gururangan, Smith (2023), Time is Encoded in the Weights of Finetuned Language Modelsarxiv.org