# FrontierScience-Research ![Category](https://img.shields.io/badge/Category-Science%20Reasoning-1f6feb) ![Metric](https://img.shields.io/badge/Metric-Rubric%20(0--10)-2da44e) ![Tasks](https://img.shields.io/badge/Tasks-60-8957e5) ![Saturation](https://img.shields.io/badge/Saturation-Low-d29922) [![Paper](https://img.shields.io/badge/Paper-arXiv-b31b1b)](https://arxiv.org/abs/2601.21165) > **Quick answer:** FrontierScience-Research is the open-ended research track of OpenAI's FrontierScience benchmark — PhD-level physics, chemistry, and biology sub-problems written and verified by domain scientists, each graded against a 10-point expert rubric (7+ points = success). It's the harder, far-from-saturated companion to FrontierScience-Olympiad. GPT-5.2 leads at 25% as of OpenAI's initial evaluations; Apodex 1.1 reports 63.3% under its own Agent Team harness. ## At a Glance **What it tests:** A model's ability to work through open-ended, PhD-level scientific research sub-problems — the kind of task a doctoral candidate, postdoc, or professor might encounter mid-research — rather than closed-form textbook questions. **Why it matters:** Prior science benchmarks (GPQA, MMLU, ScienceQA) are structured multiple-choice knowledge tests that frontier models have nearly saturated. FrontierScience-Research instead uses open-ended rubric grading to measure the harder, messier skill of sustained scientific reasoning — and remains far from saturated (best reported score 25% in OpenAI's own initial sweep, without agentic scaffolding). **Known limitations:** Grading relies on an LLM judge (GPT-5 at high reasoning effort) applying the rubric rather than fully deterministic checking, and the open-sourced gold set (60 questions) is a small held-out sample from a larger internal pool — OpenAI retains additional held-out questions specifically to track contamination. ## What FrontierScience-Research Measures FrontierScience-Research is one of two tracks in OpenAI's FrontierScience benchmark (the other, FrontierScience-Olympiad, tests closed-form olympiad-style problems). The Research track was built with 45 domain experts — postdocs, professors, and doctoral candidates — who each wrote original sub-problems representative of tasks encountered during real research in physics, chemistry, or biology, calibrated to take a qualified scientist 3–5 hours to solve. Every problem went through at least two independent expert reviews (plus a meta-review) before inclusion, and problems were explicitly filtered against internal OpenAI models — any question a model already answered well was discarded or revised, biasing the set against models similar to those used in filtering. Each Research question ships with an expert-designed rubric totaling 10 points, where each rubric item is an independently checkable pass/fail condition (e.g., "writes the following equation") covering both the final answer and the correctness of intermediate reasoning steps. OpenAI grades responses with a GPT-5 (high reasoning effort) judge model applying the rubric, averaging scores across 30 independent trials per question; a response is marked successful at 7 or more of the 10 points. The gold set — 60 questions evenly split across physics, chemistry, and biology — is open-sourced, drawn from a larger internal pool of 200+ Research questions used to guard against contamination. ## Benchmark Specifications | Field | Value | |---|---| | Task category | Reasoning / expert-level science | | Metric | Rubric score, 0–10 points; ≥7 points = success | | Number of tasks (open gold set) | 60 (evenly split: physics, chemistry, biology) | | Trials per question | 30 (averaged) | | Saturation | Low | | Created by | Wang, Lin, Hu, Jiao, Chowdhury, Chang, Patwardhan (OpenAI) | | Paper | [arXiv:2601.21165](https://arxiv.org/abs/2601.21165) | | Companion track | FrontierScience-Olympiad (closed-form olympiad problems, not tracked separately on Benchgen) | ## How FrontierScience-Research Is Scored Each response is judged by a GPT-5 model (high reasoning effort) against the question's 10-point expert rubric, which decomposes credit across both the final answer and intermediate reasoning steps. The judge returns a total point count (can be fractional); a score of 7 or more out of 10 counts as a successful solution. Reported benchmark scores are the % of questions solved successfully, averaged over 30 independent trials per question to reduce judge/sampling variance. ## State-of-the-Art Results | Rank | Model | Score | Source | Date | |---|---|---|---|---| | 1 | [GPT-5.2](/models/openai/gpt-5-2) | 25% | [FrontierScience paper (OpenAI)](https://arxiv.org/abs/2601.21165) | 2026-01 | | 2 | [Apodex 1.1](/models/apodex/apodex-1-1) | 63.3% | [Apodex 1.1 technical report](https://arxiv.org/abs/2608.23283) | 2026-08 | | 3 | [Apodex 1.1 Mini](/models/apodex/apodex-1-1-mini) | 51.7% | [Apodex 1.1 technical report](https://arxiv.org/abs/2608.23283) | 2026-08 | *The GPT-5.2 score is from OpenAI's own initial evaluation sweep without agentic scaffolding (single-pass reasoning). Apodex's higher reported scores use its own Agent Team multi-agent coordination harness rather than a single-pass setting — the two are not directly comparable methodologies, both are shown here for context. Benchgen has not independently re-run these evaluations.* ## FrontierScience-Research on Benchgen > No Benchgen results yet — [be the first to run FrontierScience-Research](/benchmarks/openai/frontierscience-research). ## FrontierScience-Research vs Other Benchmarks | Benchmark | What it tests | Tasks | Saturation | |---|---|---|---| | [GPQA Diamond](/benchmarks/rein-et-al/gpqa-diamond) | Graduate-level multiple-choice science Q&A | 198 | Medium-High | | [Humanity's Last Exam](/benchmarks/center-for-ai-safety-scale-ai/humanitys-last-exam) | Expert-level multidisciplinary knowledge | 3,000 | Low | | [CritPt](/benchmarks/critpt) | PhD-level physics research checkpoints | — | Low | FrontierScience-Research trades GPQA's fully-verifiable multiple-choice format for open-ended, rubric-graded research sub-problems — a harder and less saturated test of whether a model can actually contribute to a research workflow, not just answer a well-posed question about one. ## Run FrontierScience-Research on Your Model Benchgen lets teams run FrontierScience-Research against their own model versions, compare rubric-level performance across releases, and track regressions in scientific-reasoning capability — rather than relying on a single vendor-reported number. ## Related Models & Benchmarks - Benchmarks: [GPQA Diamond](/benchmarks/rein-et-al/gpqa-diamond) · [Humanity's Last Exam](/benchmarks/center-for-ai-safety-scale-ai/humanitys-last-exam) · [CritPt](/benchmarks/critpt) - Models: [Apodex 1.1](/models/apodex/apodex-1-1) · [Apodex 1.1 Mini](/models/apodex/apodex-1-1-mini) · [GPT-5.2](/models/openai/gpt-5-2) ## Frequently Asked Questions

What is FrontierScience-Research?

The open-ended research track of OpenAI's FrontierScience benchmark — 60 open-sourced PhD-level physics, chemistry, and biology sub-problems, each graded against a 10-point expert rubric.

What does a good FrontierScience-Research score look like?

In OpenAI's own initial sweep without agentic scaffolding, the best model (GPT-5.2) scored 25%. Higher scores from agentic systems (e.g. Apodex 1.1's 63.3%) use multi-agent coordination harnesses and aren't directly comparable to that single-pass baseline.

Who created FrontierScience-Research?

A team at OpenAI (Miles Wang, Robi Lin, Kat Hu, Joy Jiao, Neil Chowdhury, Ethan Chang, and Tejal Patwardhan), working with 45 postdoc/professor/doctoral-candidate domain experts who authored the problems.

--- *Specs from the FrontierScience paper (arXiv:2601.21165, OpenAI, Jan 2026). Last updated 2026-08-31.*

Rank	Model	Score
1	apodex-1-1	63.3
2	apodex-1-1-mini	51.7

FrontierScience-Research — Results

FrontierScience-Research — Evaluations