Claude Opus · BrowseComp · 1,266 题;单代理与多代理按协议区分 · 历代记录

按发布时间查看 Claude Opus 的 5 次发布在 BrowseComp · 1,266 题;单代理与多代理按协议区分 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 4.5

· 本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)

Claude Opus 4.5 · BrowseComp 成绩与配置
指标与成绩配置与评测条件出处
Accuracy67.8%原文 67.8% · 单位 %
公布 未确认
BrowseComp · comparison article · no thinking
设置
Single agent; web search/fetch; programmatic tool calling; context compaction at 50K; 10M total-token limit; max reasoning effort; thinking disabled.
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
BrowseComp; single-agent; web search/fetch and programmatic tool calling.
条件
50K context compaction; 10M total-token limit; max reasoning effort with thinking disabled. Kept distinct from the custom BrowseComp-Plus evaluation. 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 4 条成绩记录,含 4 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 · BrowseComp 成绩与配置
指标与成绩配置与评测条件出处
Accuracy83.73%原文 83.73% · 单位 %
公布 未确认
BrowseComp · 10M total-token limit
设置
Single agent;web search/fetch;programmatic tool calling;context compaction at 50K;10M total-token limit;Max reasoning effort;thinking disabled。
未知项
未另记
工具
允许工具
推理预算
max
框架
未记录
协议
BrowseComp;1,266 questions;single-agent;web search/fetch、programmatic tool calling、context compaction。
条件
March 6 card correction; highest single-agent config, no thinking, max reasoning effort, compaction every 50K, 10M total-token cap. Main release table’s 84.0% is the rounded value; do not merge the separately evaluated 86.57% multi-agent result.
在排名中查看此成绩
Accuracy75.83%原文 75.83% · 单位 %
公布 未确认
BrowseComp · 1M total-token limit
设置
Single agent;web search/fetch;programmatic tool calling;context compaction at 50K;1M total-token limit;Max reasoning effort;thinking disabled。
未知项
未另记
工具
允许工具
推理预算
max
框架
未记录
协议
BrowseComp;single-agent;1M total-token cap;no thinking。
条件
web search/fetch、PTC、50K context compaction、max reasoning effort;图 2.21.1.1.A。与 multi-agent result 分列。
在排名中查看此成绩
Accuracy79.9%原文 79.90% · 单位 %
公布 未确认
BrowseComp · 3M total-token limit
设置
Single agent;web search/fetch;programmatic tool calling;context compaction at 50K;3M total-token limit;Max reasoning effort;thinking disabled。
未知项
未另记
工具
允许工具
推理预算
max
框架
未记录
协议
BrowseComp;single-agent;3M total-token cap;no thinking。
条件
web search/fetch、PTC、50K context compaction、max reasoning effort;图 2.21.1.1.A。与 multi-agent result 分列。
在排名中查看此成绩
Accuracy86.57%原文 86.57% · 单位 %
公布 未确认
BrowseComp · multi-agent
设置
Orchestrator delegates to subagents; each subagent has 200K context; web search/fetch and programmatic tool calling; orchestrator compaction at 50K up to 1M total tokens; Max effort; thinking disabled。
未知项
与单代理基线不同,不作普通 effort 映射。
工具
允许工具
推理预算
max
框架
未记录
协议
BrowseComp;multi-agent;orchestrator/subagents。
条件
March 6 correction; 200K subagent context; web search/fetch, PTC; max effort, no thinking; 86.57 is not combined with single-agent results.。

Claude Opus 4.7

· 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 4.7 · BrowseComp 成绩与配置
指标与成绩配置与评测条件出处
Accuracy79.3%原文 79.3% · 单位 %
公布 未确认
发布表报告值
设置
System Card §8.1 standard configuration: adaptive thinking at max effort; default temperature/top_p; averaged over 5 trials. Context sizes are evaluation-dependent and do not exceed 1M tokens; retain all section-specific exceptions on their results.
未知项
Per-benchmark tools, context/token cap, resolution, trial count, harness and other exceptions are recorded on each result when stated.
工具
允许工具
推理预算
max
框架
未记录
协议
BrowseComp; web search, web fetch, programmatic tool calling and code execution.
条件
Thinking disabled; max effort; 10M total-token limit; context compaction triggered at 200K.
在排名中查看此成绩

Claude Opus 4.8

· 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 4.8 · BrowseComp 成绩与配置
指标与成绩配置与评测条件出处
Accuracy84.3%原文 84.3% · 单位 %
公布 未确认
BrowseComp 10M
设置
Web search、web fetch、programmatic tool calling、code execution;adaptive thinking max effort;10M token 总预算,200K 触发 context compaction;使用更新的污染规避 blocklist。
未知项
未另记
工具
允许工具
推理预算
max
框架
未记录
协议
Single-agent BrowseComp; updated contamination blocklist; search-result and fetched-page content filter also described in Appendix 9.3.
条件
Web search/fetch, programmatic tool calling and code execution; adaptive thinking/max effort; 10M token limit, compaction triggered at 200K. Exclude separate multi-agent result.
在排名中查看此成绩

Claude Opus 5

· 本次发布有 9 条成绩记录,含 9 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 · BrowseComp 成绩与配置
指标与成绩配置与评测条件出处
Accuracy90.8%原文 90.8% · 单位 %
公布 未确认
BrowseComp · max effort
设置
Single-agent BrowseComp with web search, web fetch, programmatic tool calling, and code execution; context compaction triggered at 200K; Opus 4.7 grader; card-wide default max effort.
未知项
未另记
工具
允许工具
推理预算
max
框架
未记录
协议
BrowseComp single-agent with web search, web fetch, programmatic tool calling, and code execution.
条件
90.8%; context compaction triggered at 200K tokens; Opus 4.7 grader; keep separate from the unreleased-effort budget sweep and the 10M fixed-budget effort sweep. Table 8.1.A additionally states default sampling settings, five trials, and evaluation-dependent context windows not exceeding 1M tokens; it does not state this row's per-task token budget.
在排名中查看此成绩
Accuracy86.8%原文 86.8% · 单位 %
公布 未确认
BrowseComp · token-budget sweep, unreleased effort
设置
Single-agent BrowseComp with web search, web fetch, programmatic tool calling, and code execution; context compaction triggered at 200K; per-task token budgets are recorded in each result protocol; the figure explicitly says the effort configuration was unreleased.
未知项
The effort level is not disclosed and is not inferred from token budget.
工具
允许工具
推理预算
未记录
框架
未记录
协议
BrowseComp single-agent; 1M-token task budget.
条件
The figure states Opus 5 used an unreleased effort configuration; no effort level is inferred from the budget.
Accuracy89.8%原文 89.8% · 单位 %
公布 未确认
BrowseComp · token-budget sweep, unreleased effort
设置
Single-agent BrowseComp with web search, web fetch, programmatic tool calling, and code execution; context compaction triggered at 200K; per-task token budgets are recorded in each result protocol; the figure explicitly says the effort configuration was unreleased.
未知项
The effort level is not disclosed and is not inferred from token budget.
工具
允许工具
推理预算
未记录
框架
未记录
协议
BrowseComp single-agent; 3M-token task budget.
条件
The figure states Opus 5 used an unreleased effort configuration; no effort level is inferred from the budget.
Accuracy90.2%原文 90.2% · 单位 %
公布 未确认
BrowseComp · token-budget sweep, unreleased effort
设置
Single-agent BrowseComp with web search, web fetch, programmatic tool calling, and code execution; context compaction triggered at 200K; per-task token budgets are recorded in each result protocol; the figure explicitly says the effort configuration was unreleased.
未知项
The effort level is not disclosed and is not inferred from token budget.
工具
允许工具
推理预算
未记录
框架
未记录
协议
BrowseComp single-agent; 10M-token task budget.
条件
The figure states Opus 5 used an unreleased effort configuration; no effort level is inferred from the budget.
Accuracy84.4%原文 84.4% · 单位 %
公布 未确认
BrowseComp · 10M-token budget · Low effort
设置
Single-agent BrowseComp with web search, web fetch, programmatic tool calling, and code execution; context compaction triggered at 200K; fixed 10M-token budget; figure-labeled effort; billed API costs. 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
BrowseComp single-agent; 10M-token task budget; effort scaling.
条件
Explicit effort labels from Figure 8.10.2.B; web tools and 200K-triggered context compaction described in §8.10.2.
在排名中查看此成绩
Accuracy88.9%原文 88.9% · 单位 %
公布 未确认
BrowseComp · 10M-token budget · Medium effort
设置
Single-agent BrowseComp with web search, web fetch, programmatic tool calling, and code execution; context compaction triggered at 200K; fixed 10M-token budget; figure-labeled effort; billed API costs. 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
BrowseComp single-agent; 10M-token task budget; effort scaling.
条件
Explicit effort labels from Figure 8.10.2.B; web tools and 200K-triggered context compaction described in §8.10.2.
在排名中查看此成绩
Accuracy90.2%原文 90.2% · 单位 %
公布 未确认
BrowseComp · 10M-token budget · High effort
设置
Single-agent BrowseComp with web search, web fetch, programmatic tool calling, and code execution; context compaction triggered at 200K; fixed 10M-token budget; figure-labeled effort; billed API costs. 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
BrowseComp single-agent; 10M-token task budget; effort scaling.
条件
Explicit effort labels from Figure 8.10.2.B; web tools and 200K-triggered context compaction described in §8.10.2.
在排名中查看此成绩
Accuracy90.7%原文 90.7% · 单位 %
公布 未确认
BrowseComp · 10M-token budget · Xhigh effort
设置
Single-agent BrowseComp with web search, web fetch, programmatic tool calling, and code execution; context compaction triggered at 200K; fixed 10M-token budget; figure-labeled effort; billed API costs. 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
BrowseComp single-agent; 10M-token task budget; effort scaling.
条件
Explicit effort labels from Figure 8.10.2.B; web tools and 200K-triggered context compaction described in §8.10.2.
在排名中查看此成绩
Accuracy90.8%原文 90.8% · 单位 %
公布 未确认
BrowseComp · 10M-token budget · Max effort
设置
Single-agent BrowseComp with web search, web fetch, programmatic tool calling, and code execution; context compaction triggered at 200K; fixed 10M-token budget; figure-labeled effort; billed API costs. 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
BrowseComp single-agent; 10M-token task budget; effort scaling.
条件
Explicit effort labels from Figure 8.10.2.B; web tools and 200K-triggered context compaction described in §8.10.2.
在排名中查看此成绩