Can AI truly revolutionize scientific research? The answer lies in its ability to reason deeply, not just recall facts. While AI models have made headlines for acing math and programming competitions, the real test is their capacity to contribute meaningfully to scientific discovery. But here's where it gets controversial: can AI ever truly match human intuition and creativity in the lab?
Over the past year, AI has achieved remarkable milestones, from winning gold at the International Math Olympiad to powering tools like GPT-5 that accelerate scientific workflows. Researchers are leveraging these models for tasks like cross-disciplinary literature searches and solving complex mathematical proofs, often reducing days of work to mere hours. Our paper, Early Science Acceleration Experiments with GPT-5 (https://openai.com/index/accelerating-science-gpt-5/), released in November 2025, provides early evidence of this transformative potential.
Enter FrontierScience, a new benchmark designed to measure expert-level scientific reasoning. Why? Because existing benchmarks often fall short, focusing on multiple-choice questions or lacking the depth needed to evaluate AI’s potential in real-world research. FrontierScience, crafted by experts in physics, chemistry, and biology, includes two tracks: Olympiad, for problem-solving akin to international competitions, and Research, for assessing real-world scientific capabilities. And this is the part most people miss: it’s not just about getting the right answer, but understanding the reasoning behind it.
In our initial tests, GPT-5.2 outperformed other models, scoring 77% on Olympiad and 25% on Research. While impressive, these results highlight both progress and limitations. Current models excel at structured reasoning but struggle with open-ended tasks—a stark reminder that AI still relies on human judgment for problem framing and validation. This aligns with how scientists use AI today: as a tool to accelerate workflows and explore connections, not as a replacement for human insight.
But here’s the bold question: Can AI ever independently generate groundbreaking hypotheses or interact with real-world experimental systems? FrontierScience, while a significant step forward, doesn’t fully capture these aspects. It’s a narrow lens, focusing on expert-written problems rather than the messy, multifaceted nature of scientific discovery. Yet, it provides a critical starting point, offering a standardized way to test and improve AI’s reasoning abilities.
FrontierScience consists of over 700 questions, with 100 Olympiad-style problems designed by international medalists and 60 research subtasks created by PhD scientists. These tasks were rigorously reviewed and revised to ensure they challenge even the most advanced models. We’ve open-sourced the gold sets to encourage transparency and prevent contamination. Grading is done using a rubric-based system, assessing not just the final answer but the reasoning steps—a nuanced approach that allows for detailed performance analysis.
In our evaluations, GPT-5.2 led the pack, though models like Gemini 3 Pro were close behind. While progress is undeniable, failures often stem from reasoning errors, misunderstandings of niche concepts, or factual inaccuracies. This underscores the need for continued improvement, particularly in handling open-ended tasks and real-world applications.
So, what’s next? FrontierScience is just one piece of the puzzle. As AI evolves, we’ll need more comprehensive benchmarks that evaluate hypothesis generation, multimodal interactions, and real-world experimentation. The ultimate goal? To make AI a reliable partner in scientific discovery, not just a tool for speeding up tasks. But the question remains: Can AI ever truly think like a scientist, or will it always be a step behind human ingenuity? Let us know your thoughts in the comments—we’d love to hear your take on this debate!
