OpenAI Launches FrontierScience to Test AI’s Scientific Reasoning

OpenAI has introduced FrontierScience, a new benchmark to evaluate AI’s expert-level reasoning in physics, chemistry, and biology. The benchmark measures AI’s ability to solve complex scientific problems, generate hypotheses, and support real-world research.

OpenAI Launches FrontierScience to Test AI’s Scientific Reasoning

AI Models in Research

Advanced AI models, including GPT-5, are being used for literature searches, multilingual reviews, and mathematical proofs, accelerating tasks that previously took days or weeks.

Benchmark Tracks

FrontierScience includes 700+ questions in two tracks:

• Olympiad: 100 short-answer questions designed by science olympiad medalists.

• Research: 60 PhD-level research tasks, graded with a 10-point rubric.

Initial AI Performance

• GPT-5.2: 77% on Olympiad, 25% on Research

• Gemini 3 Pro: 76% on Olympiad

Conclusion

FrontierScience helps measure AI’s scientific reasoning and problem-solving skills, showing strong progress while highlighting room for improvement in open-ended research.