Compare AI models through Arena preference charts, capability benchmarks, cost trade-offs, rankings and original evaluation evidence.
Reports, evaluations & safety observations Report statistics, measured evaluations and reported incidents retain their own meaning.
AI Incident Database · 1,801
Stanford AI Index · 2
AILuminate · 32
Hugging Face Evaluations · 2,078
LiveCodeBench · 35
τ-bench · 85
Terminal-Bench · 61
ClockQA · exact match AI Index 2026 · 2. Technical Performance · exact match · percent
Saxena et al., 2025
Qwen2-VL-7B0% 0% Llama 3.2-Vision3.23% 3.23% MiniCPM-V-2.63.23% 3.23% GPT-o14.84% 4.84% Claude-3.5-sonnet6.45% 6.45% GPT-4o8.06% 8.06% Gemini-2.022.58% 22.58% Official report and methodology ↗ Technical performance relative to human baselines AI Index 2026 · 2. Technical Performance · performance relative to human baseline · relative to human baseline
AI Index, 2026
2012 2025 Human baseline = 1 AIME GPQA Diamond ImageNet Top-5 MATH MMLU MMMU OSWorld SQuAD 2.0 SWE-bench Verified SuperGLUE VQA AIME GPQA Diamond ImageNet Top-5 MATH MMLU MMMU OSWorld SQuAD 2.0 SWE-bench Verified SuperGLUE VQA Official report and methodology ↗