AI Reasoning Model Progress May Plateau by 2026, Epoch AI Study Suggests

A study by Epoch AI, a nonprofit AI research institute, indicates that the AI sector may soon face challenges in achieving significant performance improvements from reasoning AI models. The report predicts that advancements in these models could decelerate within the next year.

Advanced reasoning models, such as OpenAI’s o3, have recently driven notable improvements in AI benchmarks, particularly in math and coding tasks. These models leverage increased computational power to enhance performance, though this often results in longer processing times compared to traditional models.

Reasoning models are created by initially training a standard model on vast datasets, followed by reinforcement learning, which provides the model with feedback to refine its problem-solving capabilities.

According to Epoch, leading AI labs like OpenAI have not yet heavily utilized computational resources for the reinforcement learning phase of reasoning model development.

This trend is shifting. OpenAI disclosed that it used approximately ten times more computational power to train o3 compared to its predecessor, o1, with Epoch suggesting that most of this was allocated to reinforcement learning. OpenAI researcher Dan Roberts recently indicated that the company plans to further prioritize reinforcement learning, potentially using even more computational resources than for initial model training.

However, Epoch notes that there is a limit to how much computational power can be applied to reinforcement learning.

Josh You, an analyst at Epoch and the study’s author, notes that performance gains from standard AI model training are currently quadrupling annually, while reinforcement learning gains are increasing tenfold every three to five months. He predicts that reasoning model progress will likely align with overall AI advancements by 2026.

Showcase at TechCrunch Sessions: AI

Reserve your space at TC Sessions: AI to present your innovations to over 1,200 decision-makers without breaking the bank. Available until May 9 or while spaces remain.

Showcase at TechCrunch Sessions: AI

Reserve your space at TC Sessions: AI to present your innovations to over 1,200 decision-makers without breaking the bank. Available until May 9 or while spaces remain.

Epoch’s study relies on certain assumptions and incorporates public statements from AI industry leaders. It also highlights that scaling reasoning models may face obstacles beyond computational limits, such as high research overhead costs.

“Ongoing research costs could limit the scalability of reasoning models,” You explains. “Since rapid computational scaling is a key factor in their progress, this warrants close attention.”

Any signs that reasoning models may hit a performance ceiling soon could raise concerns in the AI industry, which has heavily invested in their development. Research already indicates that these models, despite their high operational costs, have notable flaws, including a higher tendency to produce inaccurate outputs compared to some traditional models.