The best SWE-bench Verified score reached 98.7% of the human baseline in 2025, up from 64.3%

What the chart shows

The AI Index scales each benchmark so that the best model of each year is shown as a share of the human baseline. On SWE-bench Verified, which tests autonomous software engineering, the best score rose from 64.3% of the human baseline in 2024 to 98.7% in 2025. On OSWorld, which tests computer use, it rose from 5.7% to 75.7%.

Several reasoning benchmarks have passed the baseline: GPQA Diamond (PhD-level science) went from 47.8% in 2023 to 114.5% in 2025, and MMMU (multimodal reasoning) reached 106.1%. Older tests such as ImageNet (first passed in 2016), SQuAD 2.0 (2019) and SuperGLUE (2021) were surpassed earlier.

Key numbers

How to read it

• Each value is the best model’s score that year divided by the human baseline for that benchmark, as scaled by the AI Index; 100% means human-level performance.
• The chart shows the six benchmarks with a 2025 value. The data also covers five older ones (ImageNet, VQA, SQuAD 2.0, SuperGLUE, MMLU), all above the human baseline by their last year; download the CSV for all 11.
• The report text says SWE-bench Verified scores rose from about 60% in 2024; the data gives 64.3%, which we use.