AI evals are becoming the new compute bottleneck

AI model evaluation has become a significant compute and cost bottleneck, sometimes surpassing pretraining expenses. While early static benchmarks like HELM were already costly, newer agent-based evaluations are orders of magnitude more expensive, with single runs costing thousands of dollars. These costs are driven by complex, multi-turn rollouts, API pricing for frontier models, and the specific agent scaffolding used. Techniques that successfully reduced costs for static benchmarks, such as aggressive subsampling, are proving far less effective for these noisier and more complex agent evaluations.