The Coming Slowdown in AI’s Reasoning Revolution

Why the Breakneck Progress in Math and Coding AIs May Hit a Wall

The golden age of exponential leaps in AI reasoning may be nearing its peak. A new analysis by Epoch AI suggests performance gains for models like OpenAI’s o3—specialists in math proofs and programming—could plateau within a year. These systems, which outperform predecessors on benchmarks like MATH and HumanEval, are hitting scaling limits that even massive computing investments might not solve.

“Reinforcement learning gains grow tenfold every 3–5 months, but that curve is unsustainable,” says Epoch’s report. “We’re seeing the first signs of diminishing returns.”

Unlike standard AI training, where performance quadruples yearly, reasoning models rely on reinforcement learning—a computationally hungry process where AIs refine skills through trial and error. OpenAI poured 10x more computing power into training o3 than its predecessor, yet researchers warn that approach faces hard ceilings. The costs are staggering: each iteration requires custom reward models, human feedback pipelines, and weeks of GPU time just to validate improvements.

The Hidden Costs of Smarter AIs

Beyond hardware limits, reasoning models exhibit troubling quirks. Their knack for logical deduction comes with higher hallucination rates—confidently wrong answers that erode trust. Operational expenses also skyrocket; running an o3-style model costs 30x more than a comparable language model due to complex inference steps. “You’re trading efficiency for precision,” notes one ML engineer. “That math genius API call might burn $5 in cloud credits per question.”

OpenAI’s public roadmap doubles down on reinforcement learning, planning to allocate even greater resources to the technique. But Epoch’s data reveals a paradox: while these models solve 72% more competition-level math problems than last-gen systems, error rates on simpler arithmetic remain stubbornly high. The very training that enables breakthroughs in niche domains may leave gaping blind spots.

“We’ve entered the ‘jagged frontier’ of AI progress,” says a researcher involved in the study. “Some capabilities shoot up while others crawl—and no one knows exactly why.”

The next year will test whether reasoning models can overcome their scaling challenges or if the field needs a radical new approach. One thing’s certain: the era of easy wins is over.