Analytics from the nonprofit research institute Epoch A indicates a potential slowdown in the progress of AI models with advanced reasoning skills in the coming year. In recent months, these models, such as o3 from OpenAI, have shown significant success in solving mathematical and programming tasks, but further growth in their capabilities may not be as rapid as expected.
Particular attention is drawn to the fact that OpenAI used ten times more computational resources to train the o3 model compared to its predecessor o1 — and most of this power was directed specifically at the reinforcement learning stage. The company has already announced its intention to continue focusing on this approach and to involve even more computational resources.
Epoch analyst Josh You notes: if standard model training provides a fourfold increase in performance annually, reinforcement learning shows a tenfold increase every three to five months. However, such dynamics may quickly exhaust their resources — by two thousand twenty-six, the pace of development is likely to align with the general industry indicators.
In addition to the appetite for computational resources, reasoning models face high overhead costs for research. The research indicates: if such costs persist, scaling models may encounter unexpected limitations. It is already known that models of this class are not only expensive to use but also prone to errors, including so-called “hallucinations,” even more so than some traditional AI solutions.
