上下文窗口1.1M
最大输出128K
知识截止2025-12-01
发布日期2026-04-23
系列gpt
最后更新2026-04-23
输入模态文本图像PDF
输出模态文本
能力
推理工具调用结构化输出文件附件
基准跑分
SWE-Bench Pro 58.6 (resolve rate) · Terminal-Bench 78.2 (success rate) · SWE-Atlas Codebase QnA 45.43 (score) · SWE-Atlas Refactoring 44.79 (score) · SWE-Atlas Test Writing 42.59 (score) · Artificial Analysis Coding Agent Index 65.3 (average pass@1) · SWE-Atlas Codebase QnA 80.8 (pass@1) · SWE-Bench Pro 30.9 (pass@1) · Terminal-Bench 84.1 (pass@1) · Artificial Analysis Coding Agent Index 60.4 (average pass@1) · SWE-Atlas Codebase QnA 79.1 (pass@1) · SWE-Bench Pro 26.2 (pass@1) · Terminal-Bench 75.8 (pass@1) · Artificial Analysis Coding Agent Index 57.8 (average pass@1) · SWE-Atlas Codebase QnA 75 (pass@1) · SWE-Bench Pro 24.9 (pass@1) · Terminal-Bench 73.4 (pass@1) · Terminal-Bench 82.7 (success rate) · GPQA Diamond 93.6 (accuracy) · Humanity's Last Exam 41.4 (accuracy) · Humanity's Last Exam 52.2 (accuracy) · OSWorld-Verified 78.7 (success rate) · BrowseComp 84.4 (accuracy) · MMMU Pro 81.2 (accuracy) · ARC-AGI-2 85 (accuracy) · FrontierMath 51.7 (accuracy) · FrontierMath 35.4 (accuracy) · GDPval 84.9 (wins or ties) · MCP Atlas 75.3 (success rate) · Toolathlon 55.6 (success rate) · τ²-Bench Telecom 98 (success rate)