Claude Opus · ProgramBench · 166 个保留的 golden tasks;架构与时限按协议区分 · 历代记录

按发布时间查看 Claude Opus 的 2 次发布在 ProgramBench · 166 个保留的 golden tasks;架构与时限按协议区分 的成绩、配置与披露。

仅为直接前代间的唯一、有效精确值计算差值,要求单位、指标与明确记录的配置和条件一致。

Claude Opus 5

· 本次发布有 2 条成绩记录,含 2 条精确值(可含历史记录)

Claude Opus 5 · ProgramBench 成绩与配置
指标与成绩配置与评测条件出处
Hidden test pass rate83%原文 83% · 单位 %
公布 未确认
ProgramBench · single agent
设置
Card-wide default max effort; one agent; 166 retained golden tasks; five sequential episodes, each with a fresh context budget up to 1M tokens; no internet or decompilation tools.
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
ProgramBench; 166 golden tasks; hidden-test pass rate after episode 1.
条件
Single-agent; episode 1; each episode has a fresh context budget up to 1M tokens; max effort by card-wide default.
在排名中查看此成绩
Hidden test pass rate93%原文 93% · 单位 %
公布 未确认
ProgramBench · single agent
设置
Card-wide default max effort; one agent; 166 retained golden tasks; five sequential episodes, each with a fresh context budget up to 1M tokens; no internet or decompilation tools.
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
ProgramBench; 166 golden tasks; hidden-test pass rate after episode 5.
条件
Single-agent; five sequential episodes continue from the previous codebase; fresh context budget up to 1M tokens per episode; max effort by card-wide default.
在排名中查看此成绩

Claude Opus 5.5

· 本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)

与直接前代的记录缺少满足差值计算的同条件依据。

Claude Opus 5.5 · ProgramBench 成绩与配置
指标与成绩配置与评测条件出处
Hidden test pass rate91.2%原文 91.2% · 单位 %
公布 未确认
System Card max effort (per-result benchmark harnesses)
设置
System Card §8.1 default: adaptive thinking/max effort and default sampling settings unless a section says otherwise. Benchmark tools, evaluator, safeguards, fallback, and repetitions remain per-result conditions; this config represents the stated model effort, not harness equivalence.
未知项
Exact run time is not disclosed.、Per-benchmark tools, safeguards, and repetition counts are specified only where the source states them.
工具
工具条件未记录
推理预算
max
框架
mini-SWE-agent harness
协议
Single-agent ProgramBench; 166 retained golden tasks; mini-swe-agent harness; no six-hour time limit; no internet or decompilation.
条件
The four-task difference from the multi-agent section is not merged; score uses only tests passed by each reference binary. Attempt count is not stated. Source-wide conditions: No internet or decompilation; score only tests passed by the reference binary; do not apply the 91.2% to all 200 tasks.
在排名中查看此成绩