GPT-5 Leads as OpenAI Unveils FrontierScience: Dual‑Track Reasoning and Research Benchmark

FrontierScience Benchmark Overview

On Dec 16, 2025 OpenAI released FrontierScience, a benchmark designed to measure expert‑level scientific ability of large language models. The paper titled “FrontierScience: evaluating AI’s ability to perform expert‑level scientific tasks” reports initial scores: GPT‑5.2 achieves 25 % on the Olympiad subset and 77 % on the Research subset, surpassing other frontier models.

OpenAI states that accelerating scientific progress is a key opportunity for AI, and the benchmark aims to move beyond traditional multiple‑choice tests toward tasks that require both reasoning and research capabilities.

Dataset Design

The dataset consists of two sub‑sets:

Olympiad : created by medalists and coaches of international physics, chemistry, and biology Olympiads; focuses on short‑answer reasoning with answers that can be automatically graded (numeric, algebraic, or fuzzy‑matched biological terms).

Research : 60 original research tasks authored by PhD students, post‑docs, and professors; each task is scored on a 10‑point rubric covering modeling assumptions, reasoning steps, and intermediate conclusions.

Both subsets follow an “expert‑original + dual‑task structure + automatic scoring” design to ensure challenge, scalability, and reproducibility. From hundreds of candidates, 160 open‑source questions were selected; the rest form a hidden set for contamination detection.

Evaluation Procedure

Models were evaluated without internet access to isolate internal knowledge and reasoning. Multiple independent samplings were performed for each subset, and average scores were reported to mitigate randomness. Scoring strategies differ per subset: Olympiad uses answer equivalence with tolerance for numerical error and term matching; Research uses rubric‑based multi‑step assessment.

Results and Model Comparison

On the Olympiad subset, most frontier models scored highly. The top three were GPT‑5.2, Gemini 3 Pro, and Claude Opus 4.5, while GPT‑4o and OpenAI‑o1 lagged. This indicates that for well‑defined, closed‑form problems, current models approach expert human performance.

On the Research subset, scores dropped markedly. Errors stem from incomplete problem understanding, mishandling of key variables or assumptions, and accumulated logical mistakes in long reasoning chains. The best performers were GPT‑5, GPT‑5.2, and GPT‑5.1.

Comparing GPT‑5.2 with OpenAI‑o3 across varying token budgets shows that increasing token count raises accuracy: Olympiad accuracy rises from 67.5 % to 77.1 %, and Research from 18 % to 25 %. However, o3’s performance on Research declines at the highest token budget.

Insights and Limitations

OpenAI notes that FrontierScience does not cover the full spectrum of scientific work (e.g., experimental procedures, multimodal data, real‑world collaboration). Nonetheless, it provides a more challenging and diagnostic evaluation than saturated existing benchmarks, measuring not only answer correctness but also the ability to complete research‑style sub‑tasks.

The results suggest that large models are reliable on structured, closed‑domain scientific questions but still lack the sustained modeling and long‑chain reasoning required for authentic research tasks.