A team of 35 researchers created 50 unpublished physics problems to test how well advanced AI models can handle challenging scientific research.
According to the report, each problem would take a leading human expert approximately one day to solve. OpenAI’s GPT-6 Astra reportedly solved 49 of the 50 problems at a cost of less than $3 per problem.
AI Outperforms Researchers in a Physics Challenge
The results became even more surprising when researchers reviewed the AI’s answers.
Initially, evaluators believed several additional answers were incorrect. However, after consulting the experts who had created those problems, the team discovered mistakes in the original reference solutions. Once corrected, the AI’s answers were found to be right.
The findings raise important questions about how AI could support scientific research and complex mathematical reasoning.
Key Findings From the Experiment
- 50 unpublished problems: Researchers covered areas including quantum physics, gravity, and cosmology.
- 49 problems solved: GPT-6 Astra reportedly solved 49 out of 50 challenges for less than $3 per problem.
- Strong performance offline: GPT-6 Astra and Claude Opus 5.5 reportedly achieved scores of 92% without internet access.
- Lower-cost alternative: A cheaper DeepSeek model reportedly solved approximately half the problems.
- One unsolved challenge: None of the tested AI models reportedly solved the remaining problem.
Does This Mean AI Can Outperform Human Scientists?
The researchers suggest that some AI models may already outperform leading human experts on certain demanding scientific calculations.
However, solving a set of physics problems is not the same as conducting complete scientific research. Scientific progress also requires developing new theories, designing experiments, validating results, and reproducing findings independently.
The models’ strong offline performance may indicate that they have learned substantial amounts of physics knowledge during training, although the results alone do not establish exactly how that knowledge was acquired.
What This Means for the Future of Scientific Research
AI systems capable of solving difficult physics problems at low cost could help researchers explore mathematical approaches, test hypotheses, and accelerate complex calculations.
The next major question is whether these systems can consistently produce genuinely new scientific discoveries that withstand independent expert review.
If these results are independently validated, AI could become an increasingly powerful partner in scientific research, helping human experts tackle problems that once required days of specialized work.
