Why doesn't it know about last week?
Knowledge cutoff: training data stops at a date · Beginner · 5 min · training, context, agents · after Hallucination, RAG vs fine-tuning, Agents and MCP, System prompts and instruction priority
A knowledge cutoff is the date a model’s training data stops. It is released months later and used for years, so anything newer reaches it only through the prompt: text you paste or results from a search tool.
1 Guess first
2 Look
Where its knowledge stops
1 / 42023today · Oct 2026- training data, up to the cutoff: Aug 2024
- training and testing, then release: 5 Apr 2025 (8 months later)
- never seen: 26 months of news, up to today
model cards and vendor model pages (Meta, OpenAI, Google, Anthropic) · checked 2026-10-22
Training data stops at the cutoff. Release comes months later; today is later still.
A model is a snapshot that stops at its cutoff; news reaches it only through what is in the prompt.
The knowledge cutoff is the date the training data stops. Learning is done before release: the weights never change after that. So when you ask about last week, a model on its own can only answer from older text, and it may guess with confidence instead of saying so.
4 You are the model
You are Llama 4 Scout, trained on data up to August 2024, with no search. For three events, decide: is it in your training data or not?
you are Llama 4 Scout · training data up to Aug 2024 · no search
Meta releases Llama 3.1 · 23 Jul 2024
dates: Meta model cards, Google Gemini API docs · 2026-10-22
You are the model. Is this event in your training data?5 Check
Three habits help. Ask for its cutoff, then check the vendor’s model page: the model often just repeats its system prompt. For anything recent, paste the source into the chat. Or turn on search, so a tool fetches the page first. Either way, the news arrives in the context, not the weights.
7 Say it back
8 Mission · 5 min
With search off, ask any chatbot “What is the newest Gemini model?” Then search the same words in Exa’s dashboard playground (free monthly credit, no card), paste the top result’s text into a new chat and ask again. Did the answer move from its cutoff to this month?
Three quick ones
Drawn at random from this lesson's pool; different next time.
Done?
Go deeper
Sources
- https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct (checked 2026-10-22): Llama 4 Scout, “The pretraining data has a cutoff of August 2024”; Model Release Date April 5, 2025; sample system prompt “Your knowledge cutoff date is August 2024.”
- https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct (checked 2026-10-22): Llama 3.1 8B, “The pretraining data has a cutoff of December 2023”; Model Release Date July 23, 2024
- https://platform.claude.com/docs/en/about-claude/models/overview (checked 2026-10-22): Claude Opus 5.5 reliable knowledge cutoff Jun 2026, training data cutoff Jun 2026; “Reliable knowledge cutoff: The date through which the model’s knowledge is most extensive and reliable. Training data cutoff … is the broader range of data used.”
- https://developers.openai.com/api/docs/models/gpt-6-astra (checked 2026-10-22): GPT-6 Astra, “Apr 30, 2026 knowledge cutoff”
- https://developers.openai.com/api/docs/models/gpt-4o (checked 2026-10-22): GPT-4o, “Oct 01, 2023 knowledge cutoff”
- https://ai.google.dev/gemini-api/docs/models/gemini-2.5-flash (checked 2026-10-22): Gemini 2.5 Flash, Knowledge cutoff January 2025
- https://ai.google.dev/gemini-api/docs/models/gemini-3.8-flash (checked 2026-10-22): Latest update September 2026, no knowledge cutoff given; https://ai.google.dev/gemini-api/docs/models (last updated 2026-10-21) lists Gemini 3.8 Flash
- aifree.dev catalog, exa-free-credits (verified 2026-10-12): $10 signup credit plus $10 onboarding bonus, $10 free balance reset monthly, no payment method required
- The chat answers in the “three ways” picture are illustrative, labelled as such; the month gaps (7, 8, 34, 26) are arithmetic from the dates above and 2026-10-22
Published Oct 22, 2026 · updated Oct 22, 2026