Claude Opus 5.5

编程8 项

• ProgramBenchVals AI18.5%+
• Terminal-Bench 4.0Vals AI65.2%+
• LiveBench CodingLiveBench89.3%+
• Terminal-Bench 4.0Artificial Analysis59.6%+
• Vibe Code Bench 1–100Vals AI30.4%+
• SciCodeArtificial Analysis66.9%+
• Vibe Code Bench v1.1Vals AI90.3%+
• CyberBench Patch v1.1Vals AI85.7%+

科研9 项

• Terminal-Bench ScienceArtificial Analysis59.0%+
• Terminal-Bench ScienceVals AI47.1%+
• Mystery Game PuzzlesEpoch AI71.0%+
• FrontierMath T1–3Epoch AI91.2%+
• MysteryMechanismVals AI49.5%+
• Chess PuzzlesEpoch AI67.0%+
• Humanity's Last ExamArtificial Analysis61.4%+
• CritPtArtificial Analysis31.7%+
• LiveBench ReasoningLiveBench92.2%+

专业办公10 项

• AA-BriefcaseArtificial Analysis1,802.4+
• APEX-AgentsMercor73.5%+
• GDPval-AAArtificial Analysis1,866.3+
• τ³-BankingMercor55.0%+
• AA-Analyst AgentArtificial Analysis56.3%+
• AutomationBench-AAArtificial Analysis69.5%+
• LiveBench Data AnalysisLiveBench80.3%+
• Finance AgentVals AI58.6%+
• MedCodeVals AI49.8%+
• Harvey 法律 Agent 基准Vals AI3.8%+

知识问答7 项

• SimpleQA VerifiedEpoch AI72.2%+
• AA-OmniscienceArtificial Analysis46.4+
• MLCR-AAArtificial Analysis27.8%+
• LiveBench LanguageLiveBench86.3%+
• PBullshitBench v2BullshitBench62.0%+
• AA-LCRArtificial Analysis84.7%+
• LiveBench IFLiveBench65.7%+

真人盲选1 项

• Arena TextLMArena1,508.7+

测过，但按规则不计入

这些公开成绩未满足本榜的可比条件，具体原因如下。统一固定工具、全程使用同一型号可以采用；多模型回退、混合模型与专属异构 Agent 系统不混入模型榜。资料未说明的项目保留未知。

• Code Migration：评测方公布这次运行有 82% 的题目被转给了同厂商的旧型号，超过一半，不代表这个型号。