Claude Opus 5.5 · Toolathlon-Verified · June 2026 finalized release · 成绩与配置

Claude Opus 5.5 的 Toolathlon-Verified · June 2026 finalized release 记录。本次发布有 4 条成绩记录,含 4 条精确值(可含历史记录),附 1 条来源定位。

成绩与评测条件

Claude Opus 5.5 · Toolathlon-Verified 成绩与配置
指标与成绩配置与评测条件出处
Pass@177.8%原文 77.8% Pass@1 · 单位 %
公布 未确认
Toolathlon-Verified max effort with safeguards
设置
Adaptive thinking/max effort; production safety classifiers and sandbox-escape monitor enabled; three trials per task; all stopped trials counted as failures.
未知项
未另记
工具
工具条件未记录
推理预算
max
框架
未记录
协议
Toolathlon-Verified June 2026; 108 tasks; Pass@1 averaged over three trials per task; adaptive thinking/max effort.
条件
Production safety classifier and sandbox-escape monitor enabled; all stopped trials counted as failures. Source-wide conditions: Toolathlon-Verified June 2026; 108 tasks; three trials per task; adaptive thinking/max effort; safety classifiers enabled; one classifier stop and six sandbox-escape-monitor stops counted as failures.
在排名中查看此成绩
Pass@382.4%原文 82.4% Pass@3 · 单位 %
公布 未确认
Toolathlon-Verified max effort with safeguards
设置
Adaptive thinking/max effort; production safety classifiers and sandbox-escape monitor enabled; three trials per task; all stopped trials counted as failures.
未知项
未另记
工具
工具条件未记录
推理预算
max
框架
未记录
协议
Toolathlon-Verified June 2026; 108 tasks; at least one of three trials correct; adaptive thinking/max effort.
条件
Production safety classifier and sandbox-escape monitor enabled; all stopped trials counted as failures. Source-wide conditions: Toolathlon-Verified June 2026; 108 tasks; three trials per task; adaptive thinking/max effort; safety classifiers enabled; one classifier stop and six sandbox-escape-monitor stops counted as failures.
在排名中查看此成绩
Pass³72.2%原文 72.2% Pass³ · 单位 %
公布 未确认
Toolathlon-Verified max effort with safeguards
设置
Adaptive thinking/max effort; production safety classifiers and sandbox-escape monitor enabled; three trials per task; all stopped trials counted as failures.
未知项
未另记
工具
工具条件未记录
推理预算
max
框架
未记录
协议
Toolathlon-Verified June 2026; 108 tasks; all three trials correct; adaptive thinking/max effort.
条件
Production safety classifier and sandbox-escape monitor enabled; all stopped trials counted as failures. Source-wide conditions: Toolathlon-Verified June 2026; 108 tasks; three trials per task; adaptive thinking/max effort; safety classifiers enabled; one classifier stop and six sandbox-escape-monitor stops counted as failures.
在排名中查看此成绩
Average assistant turns26.9 turns原文 26.9 turns · 单位 turns
公布 未确认
Toolathlon-Verified max effort with safeguards
设置
Adaptive thinking/max effort; production safety classifiers and sandbox-escape monitor enabled; three trials per task; all stopped trials counted as failures.
未知项
未另记
工具
工具条件未记录
推理预算
max
框架
未记录
协议
Toolathlon-Verified June 2026; average assistant turns per trajectory across 108 tasks and three trials.
条件
Trajectory length is a separate reported metric, not a success rate. Source-wide conditions: Toolathlon-Verified June 2026; 108 tasks; three trials per task; adaptive thinking/max effort; safety classifiers enabled; one classifier stop and six sandbox-escape-monitor stops counted as failures.
在排名中查看此成绩

出处与披露

测试版本与指标

别名 Toolathlon-Verified

  • Pass@1 · % · 越高越好
  • Pass@3 · % · 越高越好
  • Pass³ · % · 越高越好
  • Average assistant turns · turns · 越低越好