Claude Sonnet 4.6 · BrowseComp(1,266 题;单代理与多代理按协议区分) 82.07% · 附件证据(ev46-browsecomp-multi-current)

Claude Sonnet 4.6 的 BrowseComp · 1,266 题;单代理与多代理按协议区分 官方出处。Adaptive thinking, max effort · 82.07%。见《Claude Sonnet 4.6 System Card》§2.20.1.3, PDF p.46。证据 ev46-browsecomp-multi-current。

原文与定位

Claude Sonnet 4.6 · BrowseComp(1,266 题;单代理与多代理按协议区分),Adaptive thinking, max effort配置,Accuracy 82.07%(成绩公布 2026-03-06)。见《Claude Sonnet 4.6 System Card》§2.20.1.3, PDF p.46。 查看成绩条件

Subagent toolset: ● Web search ● Web fetch ● Programmatic tool calling (code execution & bash) Context management: ● Subagents each get 200k context ● Context compaction for the orchestrator agent kicks in at 50k tokens , with a limit of 1M total tokens ● Effort is set to max, letting the model dynamically allocate thinking depth based on task complexity With this configuration, Claude Sonnet 4.6 achieved 82.07% accuracy, edging out the top-performing single-agent configuration by 8.1 percentage points.
官方材料
Claude Sonnet 4.6 System Card · PDF 第 46 页
定位
附件 · §2.20.1.3, PDF p.46
材料版本
PDF with changelog dated 2026-03-06; 135 pages; collected 2026-11-10
版本公布
未确认
披露时间
· 后补披露
采集时间
历史依据
暂无可靠历史版本依据
引用文本

Claude Sonnet 4.6 · BrowseComp(1,266 题;单代理与多代理按协议区分),Adaptive thinking, max effort配置,Accuracy 82.07%(成绩公布 2026-03-06)。官方出处 Claude Sonnet 4.6 System Card(附件,PDF with changelog dated 2026-03-06; 135 pages; collected 2026-11-10)§2.20.1.3, PDF p.46,https://www.anthropic.com/claude-sonnet-4-6-system-card#page=46。Benchlog https://benchlog.lekshome.cc/evidence/claude-sonnet-4-6/ev46-browsecomp-multi-current/

关联成绩与条件

Claude Sonnet 4.6 · BrowseComp 成绩与配置
指标与成绩配置与评测条件证据关系
Accuracy82.07%原文 82.07% · 单位 %
公布

更正记录 · 替代 82.62%。System Card changelog dated 2026-03-06 states 82.62% -> 82.07%; after removing leaked problems, five of eleven flagged questions remained correct.

Adaptive thinking, max effort
设置
Adaptive thinking enabled; max effort; default temperature and top_p unless an evaluation specifies otherwise.
未知项
Context window is evaluation-dependent; Table 2.1.A states it does not exceed 1M tokens.
工具
允许工具
推理预算
max
框架
未记录
协议
BrowseComp multi-agent orchestrator and subagents.
条件
Orchestrator has no direct tools and can only spawn subagents; subagents use web search, web fetch and programmatic tool calling. Orchestrator uses 200K context per subagent, 50K compaction and a 1M total-token limit; max effort. Blacklist updated and affected questions rerun.

直接支持该成绩