GPT 5-4 scores 20% on critpt, a benchmark of research-level physics problems

OpenAI's GPT-4 has achieved a score of 20% on the CritPT benchmark, a specialized evaluation consisting of research-level physics problems. This benchmark, distinct from more common general knowledge or coding tests, is designed to probe an AI's capacity for complex scientific reasoning and understanding of advanced concepts in fields like quantum mechanics and thermodynamics. The result highlights a growing focus on measuring AI capabilities in domains that could directly contribute to scientific and technological advancement, rather than just commercial or conversational utility.

The 20% score, while modest, establishes a baseline for large language models (LLMs) in tackling graduate-level physics challenges. Proponents argue that excelling in such rigorous, science-based benchmarks is critical for AI to drive real-world progress in areas like fusion energy, novel materials, and medical research. This shift in evaluation prioritizes AI's potential to augment human researchers and accelerate discovery, steering the development frontier toward tools that can expand fundamental knowledge and resource efficiency, contrasting with applications focused solely on automation or content generation.