Cognition co-founder and president Russell Kaplan said cost and speed should guide the choice of AI models for software engineering tasks they can already perform well, limiting the value of additional capability.
Kaplan wrote on Oct. 19 that many such tasks are “intelligence saturated.” His argument applies when a model meets the requirements of a particular assignment. It does not cover all software engineering work.
He said an analysis he was discussing was “directionally consistent” with Cognition’s internal data. The Information’s analysis of model capability and developer traffic identified falling prices and wider model capability as factors behind greater use of cheaper models.
Devin is Cognition’s software engineering agent. Agent performance also depends on the software around a model: in earlier EconoTimes coverage of AI agent evaluations, an evaluator said better supporting software could allow a weaker model to outperform a stronger one.
