Claude Fable 5.1 · ProgramBench(166 个保留的 golden tasks;架构与时限按协议区分) 87.6% · 附件证据(ev-f51-card-programbench)

Claude Fable 5.1 的 ProgramBench · 166 个保留的 golden tasks;架构与时限按协议区分 官方出处。Adaptive Max · 87.6%。见《Claude Fable 5.1 and Claude Mythos 5.1 System Card》本地留存副本, PDF p.176。证据 ev-f51-card-programbench。

原文与定位

Claude Fable 5.1 · ProgramBench(166 个保留的 golden tasks;架构与时限按协议区分),Adaptive Max配置,Hidden test pass rate 87.6%(具体成绩公布时间未确认)。见《Claude Fable 5.1 and Claude Mythos 5.1 System Card》本地留存副本, PDF p.176。 查看成绩条件

On this set, Claude Fable 5.1 scored 87.6% ... the same mini-swe-agent harness that the upstream repo uses, without the six-hour time limit.
官方材料
Claude Fable 5.1 and Claude Mythos 5.1 System Card · PDF 第 176 页
定位
附件 · 本地留存副本, PDF p.176
材料版本
PDF dated September 1, 2026; 212 pages. Full text review completed: pp.1-212. All 64 low-text or figure pages were rendered and visually inspected in 16 sheets; no additional standard Fable 5.1 capability scores were found in the supplementary safety/alignment sections. Revision date unknown. Local: 本地留存副本 and .txt.
版本公布
未确认
披露时间
未确认 · 首发/后补未确认
采集时间
历史依据
暂无可靠历史版本依据
引用文本

Claude Fable 5.1 · ProgramBench(166 个保留的 golden tasks;架构与时限按协议区分),Adaptive Max配置,Hidden test pass rate 87.6%(具体成绩公布时间未确认)。官方出处 Claude Fable 5.1 and Claude Mythos 5.1 System Card(附件,PDF dated September 1, 2026; 212 pages. Full text review completed: pp.1-212. All 64 low-text or figure pages were rendered and visually inspected in 16 sheets; no additional standard Fable 5.1 capability scores were found in the supplementary safety/alignment sections. Revision date unknown. Local: 本地留存副本 and .txt.)本地留存副本, PDF p.176,https://www.anthropic.com/claude-fable-5-1-mythos-5-1-system-card#page=176。Benchlog https://benchlog.lekshome.cc/evidence/claude-fable-and-mythos-5-1/ev-f51-card-programbench/

关联成绩与条件

Claude Fable 5.1 · ProgramBench 成绩与配置
指标与成绩配置与评测条件证据关系
Hidden test pass rate87.6%原文 87.6% · 单位 %
公布 未确认
Adaptive Max
设置
Claude Fable 5.1;CursorBench 主文引述明确为 max effort;后续 System Card 对 OSWorld/BenchCAD/OfficeQA 对照结果亦标 max effort。
未知项
不同 benchmark 的具体执行细节分别见来源章节;部分测试的采样参数或运行数未单独公布。
工具
工具条件未记录
推理预算
max
框架
mini-SWE-agent harness
协议
ProgramBench; 166 golden tasks after excluding 34 tasks with flaky reference binaries; same mini-swe-agent harness without its six-hour time limit.
条件
Single-agent result; multi-agent curves are nonstandard harness experiments and are not ranked.

直接支持该成绩

在排名中查看此成绩