🚀NEW COURSEVibe Coding AI Apps with Claude Code 🤖✨Enroll now
Evaluation · Reasoning

FrontierScience

Free while signed in. Answers cite the passages they came from.

Figure 1
FrontierScience
The curator’s take

OpenAI introduced FrontierScience, a new benchmark measuring AI capabilities for expert-level scientific reasoning across physics, chemistry, and biology. The benchmark consists of over 700 questions created and verified by domain experts, including international olympiad medalists and PhD scientists. - **Two evaluation tracks:** FrontierScience-Olympiad contains 100 questions designed by olympiad medalists for constrained short-answer reasoning. FrontierScience-Research has 60 open-ended research subtasks graded on 10-point rubrics. - **Benchmark results:** GPT-5.2 leads with 77% on Olympiad and 25% on Research tasks. Gemini 3 Pro scored 76% on Olympiad. The Research track shows significant room for improvement. - **Expert collaboration:** 42 former international medalists (totaling 109 olympiad medals) created Olympiad questions. 45 PhD scientists across quantum electrodynamics, synthetic chemistry, and evolutionary biology developed Research tasks. - **Why it matters:** As GPQA went from 39% with GPT-4 to 92% with GPT-5.2 in two years, FrontierScience provides harder problems to track progress toward AI-accelerated scientific discovery.

Key points
01

Two evaluation tracks: FrontierScience-Olympiad contains 100 questions designed by olympiad medalists for constrained short-answer reasoning. FrontierScience-Research has 60 open-ended research subtasks graded on 10-point rubrics.

02

Benchmark results: GPT-5.2 leads with 77% on Olympiad and 25% on Research tasks. Gemini 3 Pro scored 76% on the Olympiad. The Research track shows significant room for improvement.

03

Expert collaboration: 42 former international medalists (totaling 109 olympiad medals) created Olympiad questions. 45 PhD scientists across quantum electrodynamics, synthetic chemistry, and evolutionary biology developed Research tasks.

04

Why it matters: As GPQA went from 39% with GPT-4 to 92% with GPT-5.2 in two years, FrontierScience provides harder problems to track progress toward AI-accelerated scientific discovery.

Every Monday
Get next week’s papers.

The same picks and the same summaries, in your inbox. Free, and 176 issues deep.

Subscribe on Substack