Modeldle

Knowledge cutoffs explained

Why AI models only know the world up to a certain date, why that date lands months before release, and how web search and other tools work around it.

Last updated Oct 1, 2026

Ask an AI model about last week's news and it may confidently describe a world that has already moved on. That is not a bug in the usual sense. Every model has a knowledge cutoff: a point in time after which it saw little or no training data.

What a knowledge cutoff is

A model learns what it knows by training on a huge snapshot of text: web pages, books, code, research papers and more. That snapshot has to end somewhere, and the date where it ends is the knowledge cutoff. Anything that happened afterwards, whether an election, a product launch or a new scientific result, is missing from the model's built-in knowledge.

Labs usually publish the cutoff as a month, and Modeldle shows it on a model's page whenever the lab has published one. Some labs give two dates: when the training data ends, and an earlier date up to which they consider the model's knowledge reliable.

Why the cutoff comes months before release

A model is never released the day its data is collected. A lot has to happen in between:

  1. Preparing the data. Raw text is filtered for quality, cleaned of duplicates and spam, and checked for material that should not be there.
  2. Pretraining. The main training run can take weeks or months on thousands of chips.
  3. Post-training. The raw model is fine-tuned to follow instructions, improved with human and automated feedback, and often taught to reason step by step.
  4. Testing. Labs run benchmarks, safety evaluations and red-teaming, then fix what they find.
  5. Launch. Serving capacity, documentation and pricing all have to be ready.

In the Modeldle catalog, the result is a gap that ranges from a couple of months to more than a year:

ModelKnowledge cutoffReleasedGap
Claude Sonnet 4.5July 2025September 20252 months
Claude Opus 5.5June 2026September 20263 months
GPT-6 SolApril 2026September 20265 months
Llama 4 MaverickAugust 2024April 20258 months
GPT-5September 2024August 202511 months
Gemini 3.1 ProJanuary 2025February 202613 months
Kimi K2.6January 2025April 202615 months

The gap keeps growing after launch, too. A model that ships three months after its cutoff is a full year out of date nine months later, and people often keep using a model long after it comes out.

Newer does not always mean more up to date

You might expect each new model to know about more recent events than the last one. Usually it does, but not always. A lab may build several models from the same data snapshot, or give an existing model further training without refreshing its data. A few examples from the catalog:

  • Gemini 3.1 Pro lists the same January 2025 cutoff as Gemini 2.5 Pro, even though it came out eight months later.
  • Claude Opus 4.6 shares a May 2025 cutoff with Claude Opus 4.5.
  • Claude Haiku 4.5 was released in October 2025, a month after Claude Sonnet 4.5, yet its cutoff is February 2025, five months earlier than Sonnet's.

The fuzzy edge

A cutoff is not a clean line. The internet takes time to write about anything: first reports appear quickly, but analysis, reference articles and follow-up discussion pile up over months and years. So for the weeks just before its cutoff, a model has seen only a thin slice of what will eventually be written. Its knowledge of that period is patchier than its knowledge of earlier years.

This is also why models are often unsure of their own cutoff, and can give the wrong date when asked. They may not know much about themselves, either. A model is trained before it launches, so it has seen little or nothing about its own release, and it may not recognize newer models from its own lab.

How tools and search change the picture

A cutoff limits what a model remembers, not what it can read. Many AI apps now give models tools that bring in current information:

  • Web search. The app runs a search, places the results in the model's context window, and the model answers from them, ideally with links.
  • Documents and retrieval. You can upload files, or a company can connect its own knowledge base, so the model reads current material at the moment you ask.
  • Today's date. Many apps quietly tell the model the current date, so it can at least judge how old its own knowledge is.

This works well, with limits. The answer is only as current as what the tool finds, and a model can still blend stale memories with fresh results. When an answer depends on recent events, check whether the model actually searched, and look at its sources.

A practical rule: trust a model's built-in knowledge for things that change slowly, such as how photosynthesis works or the plot of a classic novel. Use search or your own documents for anything that changes quickly, such as prices, laws, sports results and, fittingly, AI models themselves.

In the quiz

Modeldle's date questions, like "Which came out more recently?" and "Which of these came out first?", are about release dates, not knowledge cutoffs. The two usually move together, but as the Haiku and Sonnet example shows, not always, so a model's cutoff is not a safe clue to its launch date. Model pages list both dates, so you can compare them after you play today's quiz.