Pegi News reports that OpenAI researcher Noam Brown has shared his opinion, pointing out that as the performance of AI models improves, the scores on standardized tests that measure the quality of the model are gradually shifting towards controlling the ability to infer.

A fixed and singular score no longer reflects the true level of a powerful model, and evaluation standards in the future should shift to a performance curve based on inference ability or the number of tokens generated.
As an example of the new GPT-5.5 model test, in traditional preliminary tests, there was no clear superiority of GPT-5.5 compared to GPT-5.4, but once more inference capacity was allocated, its performance began to rise explosively.
Noam Brown warned that current assessments of biological or network security often do not include a fixed inference budget, and when adversaries invest more than a million dollars at the national level in a specific task, the model that seemed safe may cross the red line of danger.