Claude Opus 5 · Frontier-Bench · v0.1 · 成绩与配置

Claude Opus 5 的 Frontier-Bench · v0.1 记录。本次发布有 5 条成绩记录,含 5 条精确值(可含历史记录),附 3 条来源定位。

成绩与评测条件

Claude Opus 5 · Frontier-Bench 成绩与配置
指标与成绩配置与评测条件出处
Score43.3%原文 43.3% · 单位 %
公布 未确认
FrontierBench v0.1 · Harbor evaluation
设置
The Table 8.1.A Harbor result; the card-wide default is adaptive thinking at max effort, default sampling, and five trials unless noted. Harness-specific details for the Harbor row are not stated.
未知项
Do not merge with the distinct Anthropic mini-SWE-agent/GKE run in §8.5.
工具
工具条件未记录
推理预算
max
框架
Harbor
协议
FrontierBench v0.1; Harbor evaluation in Table 8.1.A; distinct from the Anthropic mini-SWE-agent/GKE run in §8.5.
条件
43.3%; Table 8.1.A attributes this row to Harbor. Card-wide default max effort/default sampling/five trials applies unless noted; the Harbor row does not state its harness details.
在排名中查看此成绩
Score25%原文 25% · 单位 %
公布 未确认
FrontierBench v0.1 · mini-SWE-agent/GKE · Low effort
设置
Anthropic-run mini-SWE-agent harness on GKE; mean reward over five attempts for each of 74 unique tasks; production safety classifiers were enabled. Refusals triggered fallback to Opus 4.8 in 4% of total trials. Explicit adaptive-thinking effort: low.
未知项
The card does not state any additional sampling parameters beyond the cited evaluation description.
工具
工具条件未记录
推理预算
low
框架
mini-SWE-agent harness
协议
FrontierBench v0.1; Anthropic mini-SWE-agent harness on GKE; mean reward over 74 tasks, five attempts each.
条件
Internal harness run separate from the Harbor Table 8.1.A row; max was reported as similar to xhigh and within noise. Safety classifiers/fallback were active.
在排名中查看此成绩
Score39%原文 39% · 单位 %
公布 未确认
FrontierBench v0.1 · mini-SWE-agent/GKE · High effort
设置
Anthropic-run mini-SWE-agent harness on GKE; mean reward over five attempts for each of 74 unique tasks; production safety classifiers were enabled. Refusals triggered fallback to Opus 4.8 in 4% of total trials. Explicit adaptive-thinking effort: high.
未知项
The card does not state any additional sampling parameters beyond the cited evaluation description.
工具
工具条件未记录
推理预算
high
框架
mini-SWE-agent harness
协议
FrontierBench v0.1; Anthropic mini-SWE-agent harness on GKE; mean reward over 74 tasks, five attempts each.
条件
Internal harness run separate from the Harbor Table 8.1.A row; max was reported as similar to xhigh and within noise. Safety classifiers/fallback were active.
在排名中查看此成绩
Score44.4%原文 44.4% · 单位 %
公布 未确认
FrontierBench v0.1 · mini-SWE-agent/GKE · Xhigh effort
设置
Anthropic-run mini-SWE-agent harness on GKE; mean reward over five attempts for each of 74 unique tasks; production safety classifiers were enabled. Refusals triggered fallback to Opus 4.8 in 4% of total trials. Explicit adaptive-thinking effort: xhigh.
未知项
The card does not state any additional sampling parameters beyond the cited evaluation description.
工具
工具条件未记录
推理预算
xhigh
框架
mini-SWE-agent harness
协议
FrontierBench v0.1; Anthropic mini-SWE-agent harness on GKE; mean reward over 74 tasks, five attempts each.
条件
Internal harness run separate from the Harbor Table 8.1.A row; max was reported as similar to xhigh and within noise. Safety classifiers/fallback were active.
在排名中查看此成绩
Score43%原文 43% · 单位 %
公布 未确认
FrontierBench v0.1 · mini-SWE-agent/GKE · Max effort
设置
Anthropic-run mini-SWE-agent harness on GKE; mean reward over five attempts for each of 74 unique tasks; production safety classifiers were enabled. Refusals triggered fallback to Opus 4.8 in 4% of total trials. Explicit adaptive-thinking effort: max.
未知项
The card does not state any additional sampling parameters beyond the cited evaluation description.
工具
工具条件未记录
推理预算
max
框架
mini-SWE-agent harness
协议
FrontierBench v0.1; Anthropic mini-SWE-agent harness on GKE; mean reward over 74 tasks, five attempts each.
条件
Internal harness run separate from the Harbor Table 8.1.A row; max was reported as similar to xhigh and within noise. Safety classifiers/fallback were active.
在排名中查看此成绩

出处与披露

测试版本与指标

别名 Frontier-Bench v0.1

  • Score · % · 越高越好