Better answers aren't free
Every technique in this space, longer chains of thought, sampling multiple paths, searching over candidates, spends compute that a direct-answer model wouldn't. That compute translates into real costs: more time before a response appears, and typically more expense per query, since generating and evaluating extra reasoning tokens or candidate answers means more computation than producing one direct response.
This makes test-time compute a genuine tradeoff rather than a strictly better replacement for direct answering. For a simple factual question, spending extra compute to generate and compare five different reasoning paths adds latency and cost without meaningfully improving accuracy, because there was little chance of the direct answer being wrong in the first place.