Monday, 21 April, 2025

OpenAI’s o3 Model Scores Lower Than Expected on Independent Benchmark

OpenAI's o3 AI model, initially claimed to solve over 25% of problems on the challenging FrontierMath benchmark, has underperformed in independent evaluations. Epoch AI, the institute behind FrontierMath, reported that o3 achieved around 10% accuracy, significantly lower than OpenAI's earlier assertions. The discrepancy is attributed to differences in testing conditions, with OpenAI's internal assessments possibly utilizing more computational resources and different problem subsets.