Archive · 17 Dec 2025
OpenAI Introduces FrontierScience to Test AI's Scientific Reasoning
OpenAI has created a new test called FrontierScience to see how well AI can think like a scientist.
This test includes questions from physics, chemistry, and biology, and is designed by experts.
It has two parts: one with quick-answer questions and another with more complex research tasks.
AI models like GPT-5.2 and Gemini 3 Pro were tested, with GPT-5.2 scoring 77% on the quick-answer part and 25% on the research part.
OpenAI says this test shows that AI is getting better at scientific reasoning, but there's still room for improvement.
The goal is to use AI to help make new scientific discoveries.
OpenAI launched FrontierScience to evaluate AI's expert-level scientific reasoning across physics, chemistry, and biology.
The benchmark includes 700+ questions, with 160 in a gold-standard set, divided into Olympiad and Research tracks.
GPT-5.2 scored 77% on the Olympiad track and 25% on the Research track, while Gemini 3 Pro scored 76% on the Olympiad track.
FrontierScience focuses on reasoning, not recall, including hypothesis generation, testing, and cross-disciplinary thinking.
OpenAI acknowledges limitations, such as not assessing novel hypothesis generation or interaction with multimodal data.
- Who
- OpenAI
- What
- Launched FrontierScience benchmark to evaluate AI's expert-level scientific reasoning
- When
- December 16, 2025
- Where
- Global
- Why
- To assess AI's ability to contribute to scientific discovery and accelerate research workflows
- Benchmark Name
- FrontierScience
- Fields Covered
- Physics, Chemistry, Biology
- Total Questions
- 700+
- Gold-Standard Questions
- 160
- GPT-5.2 Olympiad Score
- 77%
- GPT-5.2 Research Score
- 25%
- Gemini 3 Pro Olympiad Score
- 76%