Does Parallel Search Use Half the Tokens?

Parag Agrawal says calling Google on every query in a deep research run gets you somewhere, but Parallel Search, for the most part, uses under half the tokens in the agent. The run becomes more accurate and faster. Using only half the tokens lets a context-limited or memory-limited model take on more problems, or do the same problems cheaper or faster.

An infinite appetite for relevant information is, at its core, an optimization problem around quality, cost, and latency. Model advances squeeze out more intelligence and then keep most of it at a tenth of the cost. The same pattern applies to search. Producing all of the signal with less noise in the results gives the model room to do more. The bet is intelligent compute allocation across the model layer, the agent layer, and the search layer.