FrontierScience-Research

FrontierScience-Research

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

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

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.

FrontierScience-Research vs Other Benchmarks

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 · Humanity's Last Exam · CritPt
• Models: Apodex 1.1 · Apodex 1.1 Mini · GPT-5.2

Frequently Asked Questions

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