Evaluating AI’s ability to perform scientific research tasks

Evaluating AI’s ability to perform scientific research tasks | OpenAI

December 16, 2025

Evaluating AI’s ability to perform scientific research tasks

We introduce FrontierScience, a new benchmark that evaluates AI capabilities for expert-level scientific reasoning across physics, chemistry, and biology.

Loading…

Share

Reasoning is at the core of scientific work. Beyond recalling facts, scientists generate hypotheses, test and refine them, and synthesize ideas across fields. As our models become more capable, the central question is how they can reason deeply to contribute to scientific research.

Over the last year, our models have reached major milestones, including achieving gold-medal performance at the International Math Olympiad and the International Olympiad in Informatics. In parallel, we’re starting to see our most capable models, such as GPT‑5, meaningfully accelerate real scientific workflows. Researchers are using these systems for tasks such as literature search across disciplines and languages and working through complex mathematical proofs. In many cases, the model shortens work that might have taken days or weeks to hours. This progress is documented in our paper Early science acceleration experiments with GPT‑5⁠, released in November 2025, which presents early evidence that GPT‑5 can measurably accelerate scientific workflows.

Introducing FrontierScience

As accelerating scientific progress is one of the most promising opportunities for AI to benefit humanity, we’re improving our models on difficult math and science tasks and working on the tools that will help scientists get the most from them.

When GPQA⁠, a “Google-Proof” science benchmark of questions written by PhD experts, was released in November 2023, GPT‑4 scored 39%, below the expert baseline of 70%. Two years later, GPT‑5.2 scored 92%. As models’ reasoning and knowledge capabilities continue to scale, more difficult benchmarks will be important to measure and forecast models’ ability to accelerate scientific research. Prior scientific benchmarks largely focus on multiple-choice questions, are saturated, or are not centrally focused on science.

To bridge this gap, we’re introducing FrontierScience: a new benchmark built to measure expert-level scientific capabilities. FrontierScience is written and verified by experts across physics, chemistry, and biology, and consists of hundreds of questions designed to be difficult, original, and meaningful. FrontierScience includes two tracks of questions: Olympiad, which measures Olympiad-style scientific reasoning capabilities, and Research, which measures real-world scientific research abilities. Providing more insight into models’ scientific capabilities helps us track progress and advance AI-accelerated science.

In our initial evaluations, GPT‑5.2 is our top performing model on FrontierScience-Olympiad (scoring 77%) and Research (scoring 25%), ahead of other frontier models. We’ve seen substantial progress on solving expert-level questions while leaving headroom for more progress, especially on open-ended research-style tasks. For scientists, this suggests that current models can already support parts of research that involve structured reasoning, while highlighting that significant work remains to improve their ability to carry out open-ended thinking. These results align with how scientists are already using today’s models: to accelerate research workflows while relying on human judgment for problem framing and validation, and increasingly to explore ideas and connections that would otherwise take much longer to uncover—including, in some cases, contributing new insights that experts then evaluate and test.

In the end, the most important benchmark for the scientific capabilities of AI is the novel discoveries it helps generate; those are what ultimately matter to science and society. FrontierScience sits upstream of that. It gives us a north star for expert-level scientific reasoning, letting us test models on a standardized set of questions, see where they succeed or fail, and identify where we need to improve them. FrontierScience is narrow and has limitations in key respects (for example, focusing on constrained, expert-written problems) and does not capture everything scientists do in their everyday work. But the field needs more difficult, original, and meaningful science benchmarks, and FrontierScience provides a step forward in this direction.

What FrontierScience measures and how we built it

The full FrontierScience evaluation spans over 700 textual questions (with 160 in the gold set) covering subfields across physics, chemistry, and biology. The benchmark is composed of an Olympiad and a Research split. FrontierScience-Olympiad contains 100 questions designed by international olympiad medalists to assess scientific reasoning in a constrained, short answer format. The Olympiad set was designed to contain theoretical questions at least as difficult as problems at international olympiad competitions. FrontierScience-Research consists of 60 original research subtasks designed by PhD scientists (doctoral candidates, professors, or postdoctoral researchers) that are graded using a 10-point rubric. The Research set was created to contain self-contained, multi-step subtasks at the level of difficulty that a PhD scientist might encounter during their research.

Sample questions

B1 reacts with aqueous bromine (Br2) to form B2. B2 reacts with potassium nitrite (KNO2) to form B3. B3 is nitrated in nitric acid (HNO3) and sulfuric acid (H2SO4) to form B4.

• B1 contains a monosubstituted aromatic 5-membered heterocycle and has a molar mass of 96.08 g/mol. It may be produced by dehydrating 5-carbon sugars (e.g. xylose) in an acid catalyst.
• B2 has the molecular formula C4H2Br2O3 and contains a tetrasubstituted alkene with 2 substituents being bromines cis to each other.
• B3 is a dipotassium salt with a molar mass of 269.27 g/mol. It contains 1 hydrogen.
• B4 is an achiral pseudohalogen dimer with 2 carbons, no hydrogens and a molar mass of 300. g/mol.

When B4 decomposes in solution, it forms an intermediate B5 and 1 equivalent of dinitrogen tetroxide (N2O4) as a side product. Intermediate B5 can be trapped and detected as a Diels-Alder adduct.

Provide the structures of B1, B2, B3, B4, and B5 in the following format, "B1: X; B2: X; B3: X; B4: X; B5: X".

Each task in FrontierScience is written and verified by a domain expert in…