Adaptive thinking; low effort, explicitly labeled in the source; benchmark-specific protocol is preserved in each result's conditions.
未知项
Sampling and trial counts vary by evaluation and are recorded per result.
工具
允许工具
推理预算
low
框架
未记录
协议
DRACO; 100 curated deep-research tasks; web search, web fetch, programmatic tool calling, code execution; 980k-token budget; five independent grader runs per response, averaged.
条件
low effort. Anthropic used an Opus 4.6 judge and a frozen offline web index; its score is not directly comparable to Perplexity’s headline score. Labeled figure value.
Adaptive thinking; medium effort, explicitly labeled in the source; benchmark-specific protocol is preserved in each result's conditions.
未知项
Sampling and trial counts vary by evaluation and are recorded per result.
工具
允许工具
推理预算
medium
框架
未记录
协议
DRACO; 100 curated deep-research tasks; web search, web fetch, programmatic tool calling, code execution; 980k-token budget; five independent grader runs per response, averaged.
条件
medium effort. Anthropic used an Opus 4.6 judge and a frozen offline web index; its score is not directly comparable to Perplexity’s headline score. Labeled figure value.
Adaptive thinking; high effort, explicitly labeled in the source; benchmark-specific protocol is preserved in each result's conditions.
未知项
Sampling and trial counts vary by evaluation and are recorded per result.
工具
允许工具
推理预算
high
框架
未记录
协议
DRACO; 100 curated deep-research tasks; web search, web fetch, programmatic tool calling, code execution; 980k-token budget; five independent grader runs per response, averaged.
条件
high effort. Anthropic used an Opus 4.6 judge and a frozen offline web index; its score is not directly comparable to Perplexity’s headline score. Labeled figure value.
Adaptive thinking; xhigh effort, explicitly labeled in the source; benchmark-specific protocol is preserved in each result's conditions.
未知项
Sampling and trial counts vary by evaluation and are recorded per result.
工具
允许工具
推理预算
xhigh
框架
未记录
协议
DRACO; 100 curated deep-research tasks; web search, web fetch, programmatic tool calling, code execution; 980k-token budget; five independent grader runs per response, averaged.
条件
xhigh effort. Anthropic used an Opus 4.6 judge and a frozen offline web index; its score is not directly comparable to Perplexity’s headline score. Labeled figure value.
Adaptive thinking; max effort, explicitly labeled in the source; benchmark-specific protocol is preserved in each result's conditions.
未知项
Sampling and trial counts vary by evaluation and are recorded per result.
工具
允许工具
推理预算
max
框架
未记录
协议
DRACO; 100 curated deep-research tasks; web search, web fetch, programmatic tool calling, code execution; 980k-token budget; five independent grader runs per response, averaged.
条件
max effort. Anthropic used an Opus 4.6 judge and a frozen offline web index; its score is not directly comparable to Perplexity’s headline score. Labeled figure value.
Official source for DRACO · 100-task Perplexity deep-research benchmark in Claude Haiku 5.5. See “Claude Haiku 5.5 System Card” at System Card PDF pp.120-121, Figure 8.8.2.A. Evidence ev-h55-card-draco.
Claude Haiku 5.5 · DRACO (100-task Perplexity deep-research benchmark), configuration System Card effort / low, Normalized score 64.3% (exact score publication date unconfirmed). See “Claude Haiku 5.5 System Card” at System Card PDF pp.120-121, Figure 8.8.2.A. View score conditions
Claude Haiku 5.5 · DRACO (100-task Perplexity deep-research benchmark), configuration System Card effort / medium, Normalized score 72.4% (exact score publication date unconfirmed). See “Claude Haiku 5.5 System Card” at System Card PDF pp.120-121, Figure 8.8.2.A. View score conditions
Claude Haiku 5.5 · DRACO (100-task Perplexity deep-research benchmark), configuration System Card effort / high, Normalized score 77.8% (exact score publication date unconfirmed). See “Claude Haiku 5.5 System Card” at System Card PDF pp.120-121, Figure 8.8.2.A. View score conditions
Claude Haiku 5.5 · DRACO (100-task Perplexity deep-research benchmark), configuration System Card effort / xhigh, Normalized score 80.5% (exact score publication date unconfirmed). See “Claude Haiku 5.5 System Card” at System Card PDF pp.120-121, Figure 8.8.2.A. View score conditions
Claude Haiku 5.5 · DRACO (100-task Perplexity deep-research benchmark), configuration System Card effort / max, Normalized score 81.5% (exact score publication date unconfirmed). See “Claude Haiku 5.5 System Card” at System Card PDF pp.120-121, Figure 8.8.2.A. View score conditions
DRACO labels Haiku 5.5 at 64.3%, 72.4%, 77.8%, 80.5%, and 81.5% for low through max effort.
Attachment · System Card PDF pp.120-121, Figure 8.8.2.A
Material version
144 页 PDF/TXT:本地留存副本、current-haiku-5-5-system-card.txt;封面日期 2026-10-07;正文 pp.1-144 已全文核读,全部63个含图/向量图表页实际视觉检查(本轮补看53页,合并此前10页),无未读页或未查看图表页;此为 current 卡片,不作为首发日附件证据。
Version published
Disclosure date
Unconfirmed · Launch or later disclosure unconfirmed
Collected
Historical basis
No reliable historical version evidence available
Citation text
Claude Haiku 5.5 · DRACO (100-task Perplexity deep-research benchmark), configuration System Card effort / low, Normalized score 64.3% (exact score publication date unconfirmed); Claude Haiku 5.5 · DRACO (100-task Perplexity deep-research benchmark), configuration System Card effort / medium, Normalized score 72.4% (exact score publication date unconfirmed); Claude Haiku 5.5 · DRACO (100-task Perplexity deep-research benchmark), configuration System Card effort / high, Normalized score 77.8% (exact score publication date unconfirmed); Claude Haiku 5.5 · DRACO (100-task Perplexity deep-research benchmark), configuration System Card effort / xhigh, Normalized score 80.5% (exact score publication date unconfirmed); Claude Haiku 5.5 · DRACO (100-task Perplexity deep-research benchmark), configuration System Card effort / max, Normalized score 81.5% (exact score publication date unconfirmed). Official source: Claude Haiku 5.5 System Card (Attachment, 144 页 PDF/TXT:本地留存副本、current-haiku-5-5-system-card.txt;封面日期 2026-10-07;正文 pp.1-144 已全文核读,全部63个含图/向量图表页实际视觉检查(本轮补看53页,合并此前10页),无未读页或未查看图表页;此为 current 卡片,不作为首发日附件证据。), System Card PDF pp.120-121, Figure 8.8.2.A, https://www.anthropic.com/claude-haiku-5-5-system-card#page=121. Benchlog https://benchlog.lekshome.cc/evidence/claude-haiku-5-5/ev-h55-card-draco/?lang=en
Related scores and conditions
Claude Haiku 5.5 · DRACO scores and configurations
Metric and score
Configuration and evaluation conditions
Evidence relation
Normalized score64.3%Reported 64.3% · Unit %
Published Unconfirmed
System Card effort / low
Settings
Adaptive thinking; low effort, explicitly labeled in the source; benchmark-specific protocol is preserved in each result's conditions.
Unknowns
Sampling and trial counts vary by evaluation and are recorded per result.
Tools
Tools allowed
Reasoning budget
low
Framework
not recorded
Protocol
DRACO; 100 curated deep-research tasks; web search, web fetch, programmatic tool calling, code execution; 980k-token budget; five independent grader runs per response, averaged.
Conditions
low effort. Anthropic used an Opus 4.6 judge and a frozen offline web index; its score is not directly comparable to Perplexity’s headline score. Labeled figure value.
Adaptive thinking; medium effort, explicitly labeled in the source; benchmark-specific protocol is preserved in each result's conditions.
Unknowns
Sampling and trial counts vary by evaluation and are recorded per result.
Tools
Tools allowed
Reasoning budget
medium
Framework
not recorded
Protocol
DRACO; 100 curated deep-research tasks; web search, web fetch, programmatic tool calling, code execution; 980k-token budget; five independent grader runs per response, averaged.
Conditions
medium effort. Anthropic used an Opus 4.6 judge and a frozen offline web index; its score is not directly comparable to Perplexity’s headline score. Labeled figure value.
Adaptive thinking; high effort, explicitly labeled in the source; benchmark-specific protocol is preserved in each result's conditions.
Unknowns
Sampling and trial counts vary by evaluation and are recorded per result.
Tools
Tools allowed
Reasoning budget
high
Framework
not recorded
Protocol
DRACO; 100 curated deep-research tasks; web search, web fetch, programmatic tool calling, code execution; 980k-token budget; five independent grader runs per response, averaged.
Conditions
high effort. Anthropic used an Opus 4.6 judge and a frozen offline web index; its score is not directly comparable to Perplexity’s headline score. Labeled figure value.
Adaptive thinking; xhigh effort, explicitly labeled in the source; benchmark-specific protocol is preserved in each result's conditions.
Unknowns
Sampling and trial counts vary by evaluation and are recorded per result.
Tools
Tools allowed
Reasoning budget
xhigh
Framework
not recorded
Protocol
DRACO; 100 curated deep-research tasks; web search, web fetch, programmatic tool calling, code execution; 980k-token budget; five independent grader runs per response, averaged.
Conditions
xhigh effort. Anthropic used an Opus 4.6 judge and a frozen offline web index; its score is not directly comparable to Perplexity’s headline score. Labeled figure value.
Adaptive thinking; max effort, explicitly labeled in the source; benchmark-specific protocol is preserved in each result's conditions.
Unknowns
Sampling and trial counts vary by evaluation and are recorded per result.
Tools
Tools allowed
Reasoning budget
max
Framework
not recorded
Protocol
DRACO; 100 curated deep-research tasks; web search, web fetch, programmatic tool calling, code execution; 980k-token budget; five independent grader runs per response, averaged.
Conditions
max effort. Anthropic used an Opus 4.6 judge and a frozen offline web index; its score is not directly comparable to Perplexity’s headline score. Labeled figure value.