An analysis from AI research institute Epoch indicates that the rapid pace of performance improvements in artificial intelligence reasoning models may slow significantly within the next year.
Reasoning models, such as OpenAI’s o3, have recently driven notable advancements in benchmarks that assess AI competencies in mathematics and programming tasks. Unlike standard models, these reasoning models approach complex tasks by applying significantly more computational power, often producing higher-quality results at the cost of longer runtimes.
Artificial intelligence researchers create reasoning models by initially training a conventional model on extensive datasets. Developers then use reinforcement learning—a technique in which the model receives iterative “feedback” based on its solutions to challenging tasks—to substantially enhance its accuracy and effectiveness.
According to Epoch, leading AI companies have not previously allocated extensive computing resources specifically to the reinforcement-learning phase of creating reasoning models. However, this is rapidly changing. OpenAI, for example, recently revealed that its newest model, o3, used approximately ten times more computing power in reinforcement learning compared to its predecessor, model o1. Moving forward, OpenAI researchers plan to continue emphasizing reinforcement learning, potentially devoting even more computational resources to this stage than to the initial training process itself.
Yet, Epoch researchers caution that there are inherent limitations to this scale-up approach. Analyst Josh You, author of the Epoch study, noted that performance enhancements from standard AI techniques are currently quadrupling annually, while progress from reinforcement learning strategies is growing at about ten-fold every several months. However, Epoch predicts that by 2026 this distinct advantage in reinforcement-learning methods may diminish, with reasoning model performance eventually converging toward broader benchmarks.
Epoch’s analysis, although reliant on certain assumptions and public statements by corporate executives, also underscores additional constraints that may hamper the continuous scaling of reasoning models. Among these are the potentially high overhead costs associated with intensive research requirements, which could limit the maximum attainable improvements. According to You, “If ongoing research consistently incurs high costs, reasoning models may not scale as far or as fast as currently anticipated. Rapid advancement in computational resources is critical to continued gains in reasoning-model performance, and this should be monitored closely.”
The conclusion that reasoning AI models could soon confront significant limitations in growth is likely to raise concern within the AI industry. Companies and laboratories have invested enormous financial and technical resources into these advanced models, even amid clear challenges such as increased operational costs and persistent accuracy issues, including a higher rate of factual “hallucinations” compared with traditional models.
