Claude Opus · Humanity's Last Exam · 2,500-question benchmark · 历代记录

按发布时间查看 Claude Opus 的 6 次发布在 Humanity's Last Exam · 2,500-question benchmark 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 4.5

· 本次发布有 3 条成绩记录,含 3 条精确值(可含历史记录)

Claude Opus 4.5 · Humanity's Last Exam 成绩与配置
指标与成绩配置与评测条件出处
Accuracy30.8%原文 30.8% · 单位 %
公布 未确认
HLE reasoning-only, no tools
设置
Reasoning-only evaluation without tools; 2,500-question HLE benchmark.
未知项
Reasoning budget, trial count and other sampling settings are not stated.
工具
禁止工具
推理预算
未记录
框架
未记录
协议
Humanity's Last Exam, 2,500 questions; reasoning-only, no tools; Claude Sonnet 4.5 grader.
条件
Reasoning budget and trial count not stated.
在排名中查看此成绩
Accuracy43.2%原文 43.2% · 单位 %
公布 未确认
HLE tools-only, no reasoning
设置
Web search, web fetch and code execution; no reasoning.
未知项
Trial count and detailed tool budgets are not stated.
工具
允许工具
推理预算
未记录
框架
未记录
协议
Humanity's Last Exam, 2,500 questions; web search, web fetch and code execution; no reasoning; Claude Sonnet 4.5 grader.
条件
Tool budgets and trial count not stated. The search-enabled arm used transcript checks and manual review to flag likely online answer retrieval.
在排名中查看此成绩
Accuracy43.4%原文 43.4% · 单位 %
公布 未确认
Humanity's Last Exam · comparison article · with tools
设置
Web search/fetch, code execution, programmatic tool calling; context compaction at 50K up to 3M total tokens; max reasoning effort; adaptive thinking enabled.
未知项
Trial count and exact sampling settings for the Opus 4.5 row are not stated.
工具
允许工具
推理预算
max
框架
未记录
协议
Humanity's Last Exam, 2,500 questions; with tools.
条件
Web search/fetch, code execution and programmatic tool calling; 50K context compaction up to 3M total tokens; adaptive thinking and max reasoning effort. Kept separate from the System Card's 43.2% tools-only/no-reasoning result. 来源主文最初发布日期为2026-02-05,但当前HTML修改于2026-10-14;没有与此数值匹配的首发图像历史字节,因此精确首次公布日未知。
在排名中查看此成绩

Claude Opus 4.6

· 本次发布有 2 条成绩记录,含 2 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 4.6 · Humanity's Last Exam 成绩与配置
指标与成绩配置与评测条件出处
Accuracy40%原文 40.0% · 单位 %
公布 未确认
HLE 无工具
设置
Humanity’s Last Exam;无工具;卡片图注明确 adaptive thinking,但未明确 effort。
未知项
effort 未说明。
工具
禁止工具
推理预算
未记录
框架
未记录
协议
Humanity's Last Exam;without tools。
条件
2,500-question Humanity’s Last Exam;reasoning only, no tools;adaptive thinking enabled;图注未明确 effort,保持 reported。
在排名中查看此成绩
Accuracy53%原文 53.0% · 单位 %
公布

raw/opus-4-6-announcement.txt line 138;官方修订说明。

HLE 有工具
设置
Max effort、adaptive thinking;web search/fetch、programmatic tool calling、code execution;50K context compaction,最多 3M total tokens。
未知项
每个配置的试验数未在系统卡段落中给出。
工具
允许工具
推理预算
max
框架
未记录
协议
Humanity's Last Exam;with tools。
条件
2,500-question HLE;web search/fetch、PTC、code execution、50K context compaction up to 3M tokens;adaptive thinking、max effort;Sonnet 4.5 grader;污染候选经 transcript review 与人工复核。System card changelog dates the score correction Feb. 17; announcement correction evidence dates it Feb. 23.
在排名中查看此成绩

Claude Opus 4.7

· 本次发布有 7 条成绩记录,含 7 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 4.7 · Humanity's Last Exam 成绩与配置
指标与成绩配置与评测条件出处
Accuracy46.9%原文 46.9% · 单位 %
公布 未确认
HLE 无工具
设置
HLE reasoning-only, no tools; thinking auto; max reasoning effort; total tokens across contexts capped at 1M; no context compaction; Claude Opus 4.6 grader; standard five-trial average applies to the headline result.
未知项
The card does not state a separate headline trial count; the §8.1 standard average of five trials applies.
工具
禁止工具
推理预算
max
框架
未记录
协议
HLE; no tools。
条件
2,500 questions; reasoning-only, no tools; thinking auto; max reasoning effort; total context tokens capped at 1M; no context compaction; Opus 4.6 grader; §8.1 five-trial average applies.
在排名中查看此成绩
Accuracy54.7%原文 54.7% · 单位 %
公布 未确认
HLE 有工具
设置
HLE with web search, web fetch, programmatic tool calling and code execution; thinking auto; max reasoning effort; total tokens across contexts capped at 1M; no context compaction; Claude Opus 4.6 grader; known HLE sources blocklisted and transcripts reviewed.
未知项
The card does not state a separate headline trial count; the §8.1 standard average of five trials applies. Single-run effort-chart points use distinct configs.
工具
允许工具
推理预算
max
框架
未记录
协议
HLE; with tools。
条件
2,500 questions; web search, web fetch, programmatic tool calling and code execution; thinking auto; max reasoning effort; 1M total-token cap; no context compaction; HLE source blocklist and Opus 4.6 transcript review; §8.1 five-trial average applies.
在排名中查看此成绩
Accuracy43%原文 43.0% · 单位 %
公布 未确认
HLE工具评测,low effort单次
设置
With tools; thinking auto; low effort; one run; total token cap 1M; no context compaction.
未知项
未另记
工具
允许工具
推理预算
low
框架
未记录
协议
HLE with web search, web fetch, programmatic tools and code execution; low effort.
条件
Single run; thinking auto; total token cap 1M; no context compaction.
在排名中查看此成绩
Accuracy48.4%原文 48.4% · 单位 %
公布 未确认
HLE工具评测,medium effort单次
设置
With tools; thinking auto; medium effort; one run; total token cap 1M; no context compaction.
未知项
未另记
工具
允许工具
推理预算
medium
框架
未记录
协议
HLE with tools; medium effort.
条件
Single run; thinking auto; total token cap 1M; no context compaction.
在排名中查看此成绩
Accuracy53.2%原文 53.2% · 单位 %
公布 未确认
HLE工具评测,high effort单次
设置
With tools; thinking auto; high effort; one run; total token cap 1M; no context compaction.
未知项
未另记
工具
允许工具
推理预算
high
框架
未记录
协议
HLE with tools; high effort.
条件
Single run; thinking auto; total token cap 1M; no context compaction.
在排名中查看此成绩
Accuracy55.4%原文 55.4% · 单位 %
公布 未确认
HLE工具评测,xhigh effort单次
设置
With tools; thinking auto; xhigh effort; one run; total token cap 1M; no context compaction.
未知项
未另记
工具
允许工具
推理预算
xhigh
框架
未记录
协议
HLE with tools; xhigh effort.
条件
Single run; thinking auto; total token cap 1M; no context compaction.
在排名中查看此成绩
Accuracy54.7%原文 54.7% · 单位 %
公布 未确认
HLE工具评测,max effort单次
设置
With tools; thinking auto; max effort; one run; total token cap 1M; no context compaction.
未知项
未另记
工具
允许工具
推理预算
max
框架
未记录
协议
HLE with tools; max effort; single run.
条件
Figure 8.8.1.B; thinking auto; 1M total-token cap; no context compaction. Keep separate from the §8.1 standard multi-trial headline result.
在排名中查看此成绩

Claude Opus 4.8

· 本次发布有 2 条成绩记录,含 2 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 4.8 · Humanity's Last Exam 成绩与配置
指标与成绩配置与评测条件出处
Accuracy49.8%原文 49.8% · 单位 %
公布 未确认
HLE 无工具
设置
Humanity’s Last Exam; reasoning only, no tools; auto thinking and max effort; 1M total-token cap; no context compaction. Some items exceed the Public API one-hour sampling limit.
未知项
具体题集与试验数未列于主表。
工具
禁止工具
推理预算
max
框架
未记录
协议
Humanity’s Last Exam; reasoning only, no tools; auto thinking, max effort; 1M total-token cap; no context compaction.
条件
The System Card says some no-tools problems exceed the Public API one-hour sampling limit; retain the reported 49.8% and source evidence with this reproducibility limitation.
Accuracy57.9%原文 57.9% · 单位 %
公布 未确认
HLE 有工具
设置
Humanity’s Last Exam; web search, web fetch, programmatic tool calling and code execution; auto thinking and max effort; 1M total-token cap; no context compaction; Opus 4.6 transcript review/regrade for confirmed answer retrieval.
未知项
具体题集与试验数未列于主表。
工具
允许工具
推理预算
max
框架
未记录
协议
Humanity’s Last Exam with web search, web fetch, programmatic tool calling and code execution; auto thinking, max effort; 1M total-token cap; no context compaction.
条件
Opus 4.6 transcript review and regrade were used to confirm HLE answer retrieval.
在排名中查看此成绩

Claude Opus 5

· 本次发布有 12 条成绩记录,含 12 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 · Humanity's Last Exam 成绩与配置
指标与成绩配置与评测条件出处
Accuracy56.3%原文 56.3% · 单位 %
公布 未确认
HLE · thinking auto, no tools
设置
Reasoning-only configuration; thinking=auto; 1M total-token cap across contexts; no context compaction; Opus 4.6 model grader.
未知项
The headline result does not identify a low-to-max effort setting.
工具
禁止工具
推理预算
max
框架
未记录
协议
Humanity’s Last Exam; reasoning-only, without tools.
条件
56.3%; thinking=auto; 1M total-token cap across contexts; no context compaction; Opus 4.6 grader. This headline run is distinct from explicit low-to-max effort charts.
在排名中查看此成绩
Accuracy64.7%原文 64.7% · 单位 %
公布 未确认
HLE · thinking auto, with tools
设置
Web search, web fetch, programmatic tool calling, and code execution; thinking=auto; 1M total-token cap across contexts; no context compaction; Opus 4.6 grader; HLE-specific URL blocklist and transcript contamination review.
未知项
The headline result does not identify a low-to-max effort setting.
工具
允许工具
推理预算
max
框架
未记录
协议
Humanity’s Last Exam; web search, web fetch, programmatic tool calling, and code execution.
条件
64.7%; thinking=auto; 1M total-token cap across contexts; no context compaction; Opus 4.6 grader; blocklist and transcript contamination review. Distinct from explicit effort charts.
在排名中查看此成绩
Accuracy56.1%原文 56.1% · 单位 %
公布 未确认
HLE effort sweep · with tools · Low effort
设置
Tools enabled; explicit figure-labeled effort scaling; billed API cost reported. This is separate from the headline thinking=auto run. Explicit adaptive-thinking effort: low.
未知项
The card does not state any additional sampling parameters beyond the cited evaluation description.
工具
允许工具
推理预算
low
框架
未记录
协议
HLE effort-scaling figure; with tools.
条件
Explicit effort labels; billed API actuals; separate from the headline thinking=auto result; figure does not state a token cap or context-compaction setting.
在排名中查看此成绩
Accuracy61.3%原文 61.3% · 单位 %
公布 未确认
HLE effort sweep · with tools · Medium effort
设置
Tools enabled; explicit figure-labeled effort scaling; billed API cost reported. This is separate from the headline thinking=auto run. Explicit adaptive-thinking effort: medium.
未知项
The card does not state any additional sampling parameters beyond the cited evaluation description.
工具
允许工具
推理预算
medium
框架
未记录
协议
HLE effort-scaling figure; with tools.
条件
Explicit effort labels; billed API actuals; separate from the headline thinking=auto result; figure does not state a token cap or context-compaction setting.
在排名中查看此成绩
Accuracy63.2%原文 63.2% · 单位 %
公布 未确认
HLE effort sweep · with tools · High effort
设置
Tools enabled; explicit figure-labeled effort scaling; billed API cost reported. This is separate from the headline thinking=auto run. Explicit adaptive-thinking effort: high.
未知项
The card does not state any additional sampling parameters beyond the cited evaluation description.
工具
允许工具
推理预算
high
框架
未记录
协议
HLE effort-scaling figure; with tools.
条件
Explicit effort labels; billed API actuals; separate from the headline thinking=auto result; figure does not state a token cap or context-compaction setting.
在排名中查看此成绩
Accuracy64.8%原文 64.8% · 单位 %
公布 未确认
HLE effort sweep · with tools · Xhigh effort
设置
Tools enabled; explicit figure-labeled effort scaling; billed API cost reported. This is separate from the headline thinking=auto run. Explicit adaptive-thinking effort: xhigh.
未知项
The card does not state any additional sampling parameters beyond the cited evaluation description.
工具
允许工具
推理预算
xhigh
框架
未记录
协议
HLE effort-scaling figure; with tools.
条件
Explicit effort labels; billed API actuals; separate from the headline thinking=auto result; figure does not state a token cap or context-compaction setting.
在排名中查看此成绩
Accuracy64.7%原文 64.7% · 单位 %
公布 未确认
HLE effort sweep · with tools · Max effort
设置
Tools enabled; explicit figure-labeled effort scaling; billed API cost reported. This is separate from the headline thinking=auto run. Explicit adaptive-thinking effort: max.
未知项
The card does not state any additional sampling parameters beyond the cited evaluation description.
工具
允许工具
推理预算
max
框架
未记录
协议
HLE effort-scaling figure; with tools.
条件
Explicit effort labels; billed API actuals; separate from the headline thinking=auto result; figure does not state a token cap or context-compaction setting.
在排名中查看此成绩
Accuracy47.8%原文 47.8% · 单位 %
公布 未确认
HLE effort sweep · no tools · Low effort
设置
No tools; explicit figure-labeled effort scaling; billed API cost reported. This is separate from the headline thinking=auto run. Explicit adaptive-thinking effort: low.
未知项
The card does not state any additional sampling parameters beyond the cited evaluation description.
工具
禁止工具
推理预算
low
框架
未记录
协议
HLE effort-scaling figure; no tools.
条件
Explicit effort labels; separate from the headline thinking=auto result; figure does not state a token cap or context-compaction setting.
在排名中查看此成绩
Accuracy54.2%原文 54.2% · 单位 %
公布 未确认
HLE effort sweep · no tools · Medium effort
设置
No tools; explicit figure-labeled effort scaling; billed API cost reported. This is separate from the headline thinking=auto run. Explicit adaptive-thinking effort: medium.
未知项
The card does not state any additional sampling parameters beyond the cited evaluation description.
工具
禁止工具
推理预算
medium
框架
未记录
协议
HLE effort-scaling figure; no tools.
条件
Explicit effort labels; separate from the headline thinking=auto result; figure does not state a token cap or context-compaction setting.
在排名中查看此成绩
Accuracy56%原文 56% · 单位 %
公布 未确认
HLE effort sweep · no tools · High effort
设置
No tools; explicit figure-labeled effort scaling; billed API cost reported. This is separate from the headline thinking=auto run. Explicit adaptive-thinking effort: high.
未知项
The card does not state any additional sampling parameters beyond the cited evaluation description.
工具
禁止工具
推理预算
high
框架
未记录
协议
HLE effort-scaling figure; no tools.
条件
Explicit effort labels; separate from the headline thinking=auto result; figure does not state a token cap or context-compaction setting.
在排名中查看此成绩
Accuracy57.6%原文 57.6% · 单位 %
公布 未确认
HLE effort sweep · no tools · Xhigh effort
设置
No tools; explicit figure-labeled effort scaling; billed API cost reported. This is separate from the headline thinking=auto run. Explicit adaptive-thinking effort: xhigh.
未知项
The card does not state any additional sampling parameters beyond the cited evaluation description.
工具
禁止工具
推理预算
xhigh
框架
未记录
协议
HLE effort-scaling figure; no tools.
条件
Explicit effort labels; separate from the headline thinking=auto result; figure does not state a token cap or context-compaction setting.
在排名中查看此成绩
Accuracy56.5%原文 56.5% · 单位 %
公布 未确认
HLE effort sweep · no tools · Max effort
设置
No tools; explicit figure-labeled effort scaling; billed API cost reported. This is separate from the headline thinking=auto run. Explicit adaptive-thinking effort: max.
未知项
The card does not state any additional sampling parameters beyond the cited evaluation description.
工具
禁止工具
推理预算
max
框架
未记录
协议
HLE effort-scaling figure; no tools.
条件
Explicit effort labels; separate from the headline thinking=auto result; figure does not state a token cap or context-compaction setting.
在排名中查看此成绩

Claude Opus 5.5

· 本次发布有 10 条成绩记录,含 10 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 · Humanity's Last Exam 成绩与配置
指标与成绩配置与评测条件出处
Accuracy67.7%原文 67.7% with tools · 单位 %
公布 未确认
HLE 有工具
设置
Humanity's Last Exam; with tools; adaptive thinking/max effort per table note。
未知项
具体题集和 trial count 未列。
工具
允许工具
推理预算
max
框架
未记录
协议
Humanity's Last Exam; with tools。
条件
发布表未列题集版本。
在排名中查看此成绩
Accuracy52.9%原文 52.9% · 单位 %
公布 未确认
HLE no tools, low effort
设置
Reasoning-only; explicit low effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader.
未知项
The figure does not state the number of model trials.
工具
禁止工具
推理预算
low
框架
未记录
协议
HLE; no tools; low effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader.
条件
Exact task-set release and model-trial count are not stated.
在排名中查看此成绩
Accuracy59%原文 59.0% · 单位 %
公布 未确认
HLE no tools, medium effort
设置
Reasoning-only; explicit medium effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader.
未知项
The figure does not state the number of model trials.
工具
禁止工具
推理预算
medium
框架
未记录
协议
HLE; no tools; medium effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader.
条件
Exact task-set release and model-trial count are not stated.
在排名中查看此成绩
Accuracy59.5%原文 59.5% · 单位 %
公布 未确认
HLE no tools, high effort
设置
Reasoning-only; explicit high effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader.
未知项
The figure does not state the number of model trials.
工具
禁止工具
推理预算
high
框架
未记录
协议
HLE; no tools; high effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader.
条件
Exact task-set release and model-trial count are not stated.
在排名中查看此成绩
Accuracy62.8%原文 62.8% · 单位 %
公布 未确认
HLE no tools, xhigh effort
设置
Reasoning-only; explicit xhigh effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader.
未知项
The figure does not state the number of model trials.
工具
禁止工具
推理预算
xhigh
框架
未记录
协议
HLE; no tools; xhigh effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader.
条件
Exact task-set release and model-trial count are not stated.
在排名中查看此成绩
Accuracy64.4%原文 64.4% · 单位 %
公布 未确认
HLE no tools, max effort
设置
Reasoning-only; explicit max effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader.
未知项
The figure does not state the number of model trials.
工具
禁止工具
推理预算
max
框架
未记录
协议
HLE; no tools; max effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader.
条件
Exact task-set release and model-trial count are not stated. Keep separate from the with-tools score.
在排名中查看此成绩
Accuracy57.4%原文 57.4% · 单位 %
公布 未确认
HLE with tools, low effort
设置
Web search, web fetch, programmatic tool calling, and code execution; explicit low effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader.
未知项
The figure does not state the number of model trials.
工具
允许工具
推理预算
low
框架
未记录
协议
HLE; web search, web fetch, programmatic tools, and code execution; low effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader.
条件
Exact task-set release and model-trial count are not stated.
在排名中查看此成绩
Accuracy63%原文 63.0% · 单位 %
公布 未确认
HLE with tools, medium effort
设置
Web search, web fetch, programmatic tool calling, and code execution; explicit medium effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader.
未知项
The figure does not state the number of model trials.
工具
允许工具
推理预算
medium
框架
未记录
协议
HLE; web search, web fetch, programmatic tools, and code execution; medium effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader.
条件
Exact task-set release and model-trial count are not stated.
在排名中查看此成绩
Accuracy63.9%原文 63.9% · 单位 %
公布 未确认
HLE with tools, high effort
设置
Web search, web fetch, programmatic tool calling, and code execution; explicit high effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader.
未知项
The figure does not state the number of model trials.
工具
允许工具
推理预算
high
框架
未记录
协议
HLE; web search, web fetch, programmatic tools, and code execution; high effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader.
条件
Exact task-set release and model-trial count are not stated.
在排名中查看此成绩
Accuracy66.4%原文 66.4% · 单位 %
公布 未确认
HLE with tools, xhigh effort
设置
Web search, web fetch, programmatic tool calling, and code execution; explicit xhigh effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader.
未知项
The figure does not state the number of model trials.
工具
允许工具
推理预算
xhigh
框架
未记录
协议
HLE; web search, web fetch, programmatic tools, and code execution; xhigh effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader.
条件
Exact task-set release and model-trial count are not stated.
在排名中查看此成绩