Via api you can turn off websearch internally. We provided all the models with their own custom tools that only provided data up to the date of the backtest.

But Grok is internally training on Tweets etc. continuously.