FrontierScience: 1,000 Expert-Level Scientific Tasks for AI Evaluation

OpenAI released this text-based benchmark dataset in December 2025 to evaluate AI performance on expert-level scientific tasks. It consists of approximately 1,000 records designed to test frontier model capabilities in complex scientific domains.

Use Cases

• Benchmarking LLM performance on expert-level scientific reasoning tasks
• Detecting benchmark contamination in training sets using the canary GUID
• Evaluating model accuracy across various scientific disciplines

Strengths

• Authored by OpenAI for frontier model evaluation
• Includes a canary GUID to prevent training data contamination
• Focuses on expert-level scientific domain tasks

Limitations

• Small sample size of approximately 1,000 records
• Lack of detailed column-level documentation in the metadata

Provenance

Source

OpenAI

Freshness

Last updated December 2025.