Knowledge cutoff
A knowledge cutoff is the latest period a model can reliably know from its training alone.
A cutoff is a property of a particular model release, not a permanent date for an entire product. Anthropic’s current model table separates a reliable knowledge cutoff from a training data cutoff, and those two dates can differ for the same model. Google’s Gemini 3 guide, for example, publishes January 2025 as the series’ knowledge cutoff.
The date describes reliability, not a hard wall. OpenAI reported that GPT-4 generally lacked knowledge of events after September 2021, and that historical example also says the model did not learn from its experience. A model can still fail on older facts, so “before the cutoff” should never be read as “known and correct.”
New facts therefore need a separate path into an answer, such as supplied documents, search results or retrieved records. Retrieval-augmented generation keeps an external, non-parametric memory alongside the model’s parametric memory. Replacing that external memory can update what the system answers as the world changes. Current evidence still needs verification because tools and sources introduce their own failure modes.
Fluent answers can sound current even when the model's built-in information is old.
Check whether a question needs information newer than the model's reliable knowledge.
- 1 · identifyNotice whether the answer depends on a changing fact or a particular date.
- 2 · compareCompare that date with the selected model's published reliable-knowledge or training-data cutoff.
- 3 · retrieveIf the fact may be newer, fetch a current source and add its relevant content to the model's context.
- 4 · answerGenerate from the supplied evidence, state its date and preserve a path back to the source.
The cutoff describes built-in knowledge, not the freshness of evidence supplied at answer time.
| Who | What they ask | What it works with |
|---|---|---|
| News assistant | “What happened in today's election count?” | A current result from the responsible election authority |
| Developer | “Does this library version support the new API?” | The versioned official documentation and release notes |
| Analyst | “Who currently leads this organisation?” | A dated official leadership page |
| Support agent | “What policy applies today?” | The current policy revision and effective date |
- A published cutoff gives users a boundary for judging when built-in knowledge may be stale.
- A freshness check helps route recent questions to search, retrieval or another live data source.
- Retrieved evidence can update an answer without retraining the model's parameters.
- Showing the evidence date makes the answer easier to audit later.
- A cutoff does not guarantee that every earlier fact appeared in training or can be recalled correctly.
- It does not make answers about earlier events automatically correct.
- Search does not make an answer automatically correct because retrieved sources can be wrong, stale or misread.
- A newer cutoff does not tell you the model's context window or maximum output.
Sources used
This explainer is written in original language. The links below support its factual claims.
- officialGPT-4, OpenAI · read 27 Sept 2026
- docsModels overview, Anthropic · read 27 Sept 2026
- docsGemini 3 developer guide, Google AI for Developers · read 27 Sept 2026
- officialAn update on our election safeguards, Anthropic · read 27 Sept 2026
- paperRetrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, Meta AI Research · read 27 Sept 2026
- officialDoes ChatGPT tell the truth?, OpenAI · read 27 Sept 2026