Knowledge Cutoff
The date after which an LLM has no training data. Events, releases, and facts from after the cutoff are unknown to the model unless retrieved at runtime.
What Is a Knowledge Cutoff
A knowledge cutoff is the last date represented in an LLM’s training corpus. Everything that happened after that date — new product launches, news, library versions, API changes — is invisible to the base model. Ask about it and you get either a refusal, an outdated answer, or, worse, a confident hallucination. Use the cutoff as a constraint when deciding whether to fine-tune, RAG, or web-search.
Cutoffs for Common Models (as of 2026)
- GPT-4o — October 2023
- GPT-4.5 / o1 / o3 — October 2023 to mid-2024 depending on variant
- Claude Sonnet 4.5 — April 2025
- Claude Opus 4.7 — January 2026
- Gemini 2.5 Pro — early 2025
- Llama 3.1 — December 2023
Providers sometimes refresh the cutoff in minor updates without changing the model name, so check the model card before trusting a number.
Cutoff vs Release Date
These are different. A model released in March 2026 might still have an October 2025 cutoff because training takes months and post-training takes more. Always compare cutoff, not release date.
How to Work Around It
- Retrieval-Augmented Generation (RAG) — pull current documents into context at query time
- Web search tools — give the model a function to fetch live pages
- System prompt facts — inject today’s date and key updates as context
- Fine-tuning — expensive and slow; rarely the right answer for freshness
If your application depends on data newer than the cutoff, RAG or tool use is mandatory.