You don’t need a frontier model. You need a verifier.

Agentic search systems can be improved more effectively with a verifier model rather than simply using a more powerful generator. Standard majority voting often fails because it can certify a popular but incorrect answer, whereas a verifier independently re-researches each candidate to veto wrong ones. This approach has been shown to surpass state-of-the-art results on the FACTS-Search benchmark by filtering the outputs of an ensemble of models. While using a frontier model as a verifier is effective but costly, training a small, specialized verifier model provides a highly cost-efficient alternative that achieves most of the accuracy gains.