If you send a top student to the International Physics Olympiad (IPhO), they might win gold. But put them in a lab to independently tackle a frontier research problem, and they might not even find the door.
AI faces the same divide between "exam hall and battlefield." OpenAI's recently released top-tier science benchmark, FrontierScience (2026), delivered a harsh report card for large language models: AI is nearly godlike at solving problems, but in the face of genuine scientific discovery, it is still a toddler learning to speak.
The Two Faces of Science: Olympiad vs. Research
OpenAI argues that judging whether an AI truly "understands science" requires looking at two things:
- Olympiad track: Extremely difficult problems covering physics, chemistry, and biology at the highest international competition level. This tests a model's core problem-solving strength—formula derivation and logical rigor.
- Research track: The real battlefield. It requires the model to act like a PhD student: forming hypotheses from ambiguous clues, designing experiments, and analyzing data no one has seen before.
- The test-taker: On the Olympiad track, it scored a high 77%, meaning most human experts no longer hold any problem-solving advantage over it.
- The research novice: Yet once it entered the Research track, its score instantly fell to 25%.
- Hypothesis generation: AI excels at running along established paths but struggles to think outside the box toward counterintuitive solutions.
- Error tolerance: Scientific research is full of failure. AI is accustomed to finding the "correct path"; when experimental data defies logic, it often hallucinates a forced explanation instead of reflecting on its underlying assumptions.
- Long-horizon planning: A scientific experiment can run for months. While AI's attention span is increasing, it remains weak at deep, multi-step decision-making across such long time scales.
A Sobering Report Card: GPT-5.2's Ceiling
Top-tier GPT-5.2's performance in this exam gives pause:
A Feynman-style analogy: imagine a student who has memorized an entire physics textbook and can instantly solve the most complex projectile equations. But drop them on a deserted island and ask them to invent a fluorescent lamp from available resources—they can do nothing but recite. That is AI's current state: it has top-tier problem-solving intuition but severely lacks scientific creativity.
Why AI Still Can't Replace Scientists
The benchmark paper points to three major weaknesses of AI in real research:
Commentary
The arrival of FrontierScience marks a new era in AI training: the leap from "imitating human conversation" to "solving humanity's unsolved problems."
A research score of 25 may look paltry, but that is precisely the most fascinating and promising uncharted territory of intelligence. When AI can cross that 50-point gap—transforming from "a machine that solves problems" into "a peer that thinks"—the pace of human civilization's progress will be rewritten.
Discussion: If AI ever truly achieves 100 on the Research track, what do you think human scientists should do? Join the ultimate debate in the comments!
*Note: This article is based on OpenAI's science evaluation benchmark released in May 2026.*