That could be mitigated by using a much smaller non-reasoning model for searches and reading. There's no reason why a search should take loads of tokens.