Live web search
Models have a training cutoff. Web search removes it: toggle the search badge under the prompt box and the model can look things up before answering.
How it works
When search is active, the model runs up to three steps behind the scenes:
- Search: find candidate pages for the query.
- Read: fetch and extract the most promising results.
- Rerank: sort the extracted content by relevance before it reaches the model.
The answer cites its sources, so you can verify every claim with one click.
When to use it
- Current events, prices, releases, anything after the model's cutoff.
- Facts you want sourced rather than recalled.
- Niche topics where training data is thin.
Leave it off for pure reasoning, writing, or coding tasks: it adds latency and cost without adding quality there.
Cost
Each pipeline step is billed against your credit like a model call, and the answer shows the total below it, search steps included. A typical searched answer costs a few cents.
Works with every model
Search is a tool the gateway provides, not a model feature. It works the same whether you are talking to Claude, GPT, Gemini, or an open-source model, and you can combine it with agents so a saved assistant always searches.