FundamentalsBeginner
Knowledge Cutoff
The date after which an AI model has no training data — and doesn't know what happened.
Every LLM has a knowledge cutoff — the date at which its training data collection ended. The model has no information about events, publications, prices, or code updates that occurred after this date.
Practical implications:
- Ask about a recent event → the model may confabulate (hallucinate) or say it doesn't know
- Ask for today's date → the model doesn't know (unless told via system prompt)
- API pricing, library versions, political situations → likely outdated
How to work around it:
- Tell the model today's date in the system prompt — it can then adjust its answers accordingly
- Use RAG to inject current documents before the model answers
- Use models with web search (Perplexity, ChatGPT with browsing, Claude with search tools)
- Build time context in — paste the relevant current information directly into the prompt
Common cutoff dates:
- Claude 3.5 Sonnet: April 2024
- GPT-4o: October 2023
- Claude Sonnet 4.6: August 2025
- Always verify at the provider's docs — cutoffs are updated with new model versions.
In plain terms
A brilliant researcher who went into isolation in April 2024 and has been completely offline since. Extraordinarily knowledgeable — but genuinely unaware of anything that happened after that date.