Accuracy of AI models in selected benchmarks, 2018-2024

Sources

IEA analysis based on data from EpochAI (2025), AI Benchmarking Hub, Papers With Code (2025), Code Generation on HumanEval and Stanford University (2024), The 2024 AI Index Report.

Notes

The y-axis refers to the accuracy of models in the benchmark test under consideration. It does not compare the ability of a model vis-à-vis a human. The benchmark test measuring coding ability is HumanEval (Code Generation). The benchmark test measuring scientific understanding is the Graduate-Level Google-Proof Q&A Benchmark (GPQA). The benchmark test measuring mathematics is the Mathematics Assessment of Textual Heuristics (MATH) Level 5. The benchmark test measuring visual understanding is Visual Commonsense Reasoning (VCR).