Knowledge Cutoff
The latest date covered by an AI model's built-in training knowledge.
Your model knows yesterday's news. Yesterday.
The knowledge cutoff problem
All models are trained on data up to a specific date, the knowledge cutoff. Claude's was trained up to early 2024. A question about a 2026 announcement gets a shrug: "I don't know."
A model with web search can look it up and give you today's answer.
What uses web search
Use search for anything current, like news, stock prices, recent announcements, and today's weather. Use it for anything niche too, like obscure research papers, job listings, and recent blog posts.
You do not need search for general knowledge like history, math, and how things work.
The cost
Web search adds latency, because the model has to query the internet, which takes time. It also adds cost, because you pay for the search plus the extra tokens it uses to read the results.
Most models let you toggle search off for simple questions and keep your bill small.
Which models have it
Claude has it via Claude.ai. ChatGPT has it with the web-search setting. Gemini has Google search built in. Some API tiers offer search too, so check the provider.