ARC-AGI v2厂商自报 · 越高越好
42.5%
Thinking mode, pass@2, public set
成绩出处CharXiv-R厂商自报 · 越高越好
86.4%
Thinking mode
成绩出处DeepSearchQA厂商自报 · 越高越好
74.8%
Thinking mode
成绩出处ERQA厂商自报 · 越高越好
64.7%
Thinking mode
成绩出处FrontierScience Research厂商自报 · 越高越好
38.3%
Contemplating mode, averaged over 8 samples
成绩出处GPQA厂商自报 · 越高越好
89.5%
Thinking mode, GPQA Diamond, averaged over 4 runs
成绩出处HealthBench Hard厂商自报 · 越高越好
42.8%
Thinking mode
成绩出处Humanity's Last Exam厂商自报 · 越高越好
58.4%
Contemplating mode, with tools (bash + browser)
成绩出处IPhO 2025厂商自报 · 越高越好
82.6%
Contemplating mode, blinded human evaluation, averaged over 3 generations
成绩出处LiveCodeBench Pro厂商自报 · 越高越好
0.8分
Thinking mode, pass@1, 2025Q2, C++, averaged over 4 attempts
成绩出处MedXpertQA厂商自报 · 越高越好
78.4%
MM
成绩出处MMMU-Pro厂商自报 · 越高越好
80.4%
Thinking mode
成绩出处ScreenSpot Pro厂商自报 · 越高越好
84.1%
Thinking mode, with Python (cropping tool)
成绩出处SimpleVQA厂商自报 · 越高越好
0.71%
Thinking mode
成绩出处SWE-Bench Pro厂商自报 · 越高越好
52.4%
Thinking mode, averaged over 4 attempts
成绩出处SWE-Bench Verified厂商自报 · 越高越好
77.4%
Thinking mode, averaged over 15 attempts
成绩出处Tau2 Telecom厂商自报 · 越高越好
91.5%
Artificial Analysis
成绩出处Terminal-Bench 2.0厂商自报 · 越高越好
59%
Thinking mode, bash-tool-only, averaged over 15 attempts
成绩出处ZEROBench厂商自报 · 越高越好
0.33%
Thinking mode, with Python, pass@5
成绩出处