Claude Opus 5.5 · DRACO(100-task Perplexity deep-research benchmark) · 附件证据(ev-op55-card-draco)
Claude Opus 5.5 的 DRACO · 100-task Perplexity deep-research benchmark 官方出处。见《Claude Opus 5.5 System Card》§8.11.2, pp.186-187, Figure 8.11.2.A。证据 ev-op55-card-draco。
Claude Opus 5.5 · DRACO(100-task Perplexity deep-research benchmark),DRACO low effort, in-house agent harness配置,Normalized score 72.5%(具体成绩公布时间未确认)。见《Claude Opus 5.5 System Card》§8.11.2, pp.186-187, Figure 8.11.2.A。 查看成绩条件
Claude Opus 5.5 · DRACO(100-task Perplexity deep-research benchmark),DRACO medium effort, in-house agent harness配置,Normalized score 83.9%(具体成绩公布时间未确认)。见《Claude Opus 5.5 System Card》§8.11.2, pp.186-187, Figure 8.11.2.A。 查看成绩条件
Claude Opus 5.5 · DRACO(100-task Perplexity deep-research benchmark),DRACO high effort, in-house agent harness配置,Normalized score 85%(具体成绩公布时间未确认)。见《Claude Opus 5.5 System Card》§8.11.2, pp.186-187, Figure 8.11.2.A。 查看成绩条件
Claude Opus 5.5 · DRACO(100-task Perplexity deep-research benchmark),DRACO xhigh effort, in-house agent harness配置,Normalized score 86.7%(具体成绩公布时间未确认)。见《Claude Opus 5.5 System Card》§8.11.2, pp.186-187, Figure 8.11.2.A。 查看成绩条件
Claude Opus 5.5 · DRACO(100-task Perplexity deep-research benchmark),DRACO max effort, in-house agent harness配置,Normalized score 87.4%(具体成绩公布时间未确认)。见《Claude Opus 5.5 System Card》§8.11.2, pp.186-187, Figure 8.11.2.A。 查看成绩条件
Normalized score on DRACO at a 980k task budget as reasoning effort rises from low to max against average cost per task.
官方当前 System Card;封面日期 2026-09-22;本地 PDF/TXT:本地留存副本、current-opus-5-5-system-card.txt;230 页全文核读;当前修订时间未知
版本公布
未确认
披露时间
未确认 · 首发/后补未确认
采集时间
历史依据
暂无可靠历史版本依据
引用文本
Claude Opus 5.5 · DRACO(100-task Perplexity deep-research benchmark),DRACO low effort, in-house agent harness配置,Normalized score 72.5%(具体成绩公布时间未确认);Claude Opus 5.5 · DRACO(100-task Perplexity deep-research benchmark),DRACO medium effort, in-house agent harness配置,Normalized score 83.9%(具体成绩公布时间未确认);Claude Opus 5.5 · DRACO(100-task Perplexity deep-research benchmark),DRACO high effort, in-house agent harness配置,Normalized score 85%(具体成绩公布时间未确认);Claude Opus 5.5 · DRACO(100-task Perplexity deep-research benchmark),DRACO xhigh effort, in-house agent harness配置,Normalized score 86.7%(具体成绩公布时间未确认);Claude Opus 5.5 · DRACO(100-task Perplexity deep-research benchmark),DRACO max effort, in-house agent harness配置,Normalized score 87.4%(具体成绩公布时间未确认)。官方出处 Claude Opus 5.5 System Card(附件,官方当前 System Card;封面日期 2026-09-22;本地 PDF/TXT:本地留存副本、current-opus-5-5-system-card.txt;230 页全文核读;当前修订时间未知)§8.11.2, pp.186-187, Figure 8.11.2.A,https://www.anthropic.com/claude-opus-5-5-system-card#page=187。Benchlog https://benchlog.lekshome.cc/evidence/claude-opus-5-5/ev-op55-card-draco/
关联成绩与条件
Claude Opus 5.5 · DRACO 成绩与配置
指标与成绩
配置与评测条件
证据关系
Normalized score72.5%原文 72.5% · 单位 %
公布 未确认
DRACO low effort, in-house agent harness
设置
Explicit low effort; web search, web fetch, programmatic tool calling, and code execution; 980k-token task budget.
未知项
Model run date and direct API reproduction are not stated.
工具
允许工具
推理预算
low
框架
未记录
协议
DRACO; low effort; web search/fetch, programmatic tools, code execution; 980k-token budget; own agent harness; five judge grading runs per response, mean; Claude Opus 4.6 grader.
条件
Judge choice can shift absolute scores 10-25 points; not directly comparable to paper headline scores. Cost is shown only as a chart axis and is not added as an inferred exact value. Source-wide conditions: 100 tasks; web search, web fetch, programmatic tool calling, code execution; five judge grading runs per response, mean reported; Claude Opus 4.6 grader; not directly comparable to paper headline scores.
Explicit medium effort; web search, web fetch, programmatic tool calling, and code execution; 980k-token task budget.
未知项
Model run date and direct API reproduction are not stated.
工具
允许工具
推理预算
medium
框架
未记录
协议
DRACO; medium effort; web search/fetch, programmatic tools, code execution; 980k-token budget; own agent harness; five judge grading runs per response, mean; Claude Opus 4.6 grader.
条件
Judge choice can shift absolute scores 10-25 points; not directly comparable to paper headline scores. Cost is not inferred from the chart. Source-wide conditions: 100 tasks; web search, web fetch, programmatic tool calling, code execution; five judge grading runs per response, mean reported; Claude Opus 4.6 grader; not directly comparable to paper headline scores.
Explicit high effort; web search, web fetch, programmatic tool calling, and code execution; 980k-token task budget.
未知项
Model run date and direct API reproduction are not stated.
工具
允许工具
推理预算
high
框架
未记录
协议
DRACO; high effort; web search/fetch, programmatic tools, code execution; 980k-token budget; own agent harness; five judge grading runs per response, mean; Claude Opus 4.6 grader.
条件
Judge choice can shift absolute scores 10-25 points; not directly comparable to paper headline scores. Cost is not inferred from the chart. Source-wide conditions: 100 tasks; web search, web fetch, programmatic tool calling, code execution; five judge grading runs per response, mean reported; Claude Opus 4.6 grader; not directly comparable to paper headline scores.
Explicit xhigh effort; web search, web fetch, programmatic tool calling, and code execution; 980k-token task budget.
未知项
Model run date and direct API reproduction are not stated.
工具
允许工具
推理预算
xhigh
框架
未记录
协议
DRACO; xhigh effort; web search/fetch, programmatic tools, code execution; 980k-token budget; own agent harness; five judge grading runs per response, mean; Claude Opus 4.6 grader.
条件
Judge choice can shift absolute scores 10-25 points; not directly comparable to paper headline scores. Cost is not inferred from the chart. Source-wide conditions: 100 tasks; web search, web fetch, programmatic tool calling, code execution; five judge grading runs per response, mean reported; Claude Opus 4.6 grader; not directly comparable to paper headline scores.
Explicit max effort; web search, web fetch, programmatic tool calling, and code execution; 980k-token task budget.
未知项
Model run date and direct API reproduction are not stated.
工具
允许工具
推理预算
max
框架
未记录
协议
DRACO; max effort; web search/fetch, programmatic tools, code execution; 980k-token budget; own agent harness; five judge grading runs per response, mean; Claude Opus 4.6 grader.
条件
Judge choice can shift absolute scores 10-25 points; not directly comparable to paper headline scores. Cost is not inferred from the chart. Source-wide conditions: 100 tasks; web search, web fetch, programmatic tool calling, code execution; five judge grading runs per response, mean reported; Claude Opus 4.6 grader; not directly comparable to paper headline scores.
Claude Opus 5.5 · DRACO (100-task Perplexity deep-research benchmark) · Attachment evidence (ev-op55-card-draco)
Official source for DRACO · 100-task Perplexity deep-research benchmark in Claude Opus 5.5. See “Claude Opus 5.5 System Card” at §8.11.2, pp.186-187, Figure 8.11.2.A. Evidence ev-op55-card-draco.
Claude Opus 5.5 · DRACO (100-task Perplexity deep-research benchmark), configuration DRACO low effort, in-house agent harness, Normalized score 72.5% (exact score publication date unconfirmed). See “Claude Opus 5.5 System Card” at §8.11.2, pp.186-187, Figure 8.11.2.A. View score conditions
Claude Opus 5.5 · DRACO (100-task Perplexity deep-research benchmark), configuration DRACO medium effort, in-house agent harness, Normalized score 83.9% (exact score publication date unconfirmed). See “Claude Opus 5.5 System Card” at §8.11.2, pp.186-187, Figure 8.11.2.A. View score conditions
Claude Opus 5.5 · DRACO (100-task Perplexity deep-research benchmark), configuration DRACO high effort, in-house agent harness, Normalized score 85% (exact score publication date unconfirmed). See “Claude Opus 5.5 System Card” at §8.11.2, pp.186-187, Figure 8.11.2.A. View score conditions
Claude Opus 5.5 · DRACO (100-task Perplexity deep-research benchmark), configuration DRACO xhigh effort, in-house agent harness, Normalized score 86.7% (exact score publication date unconfirmed). See “Claude Opus 5.5 System Card” at §8.11.2, pp.186-187, Figure 8.11.2.A. View score conditions
Claude Opus 5.5 · DRACO (100-task Perplexity deep-research benchmark), configuration DRACO max effort, in-house agent harness, Normalized score 87.4% (exact score publication date unconfirmed). See “Claude Opus 5.5 System Card” at §8.11.2, pp.186-187, Figure 8.11.2.A. View score conditions
Normalized score on DRACO at a 980k task budget as reasoning effort rises from low to max against average cost per task.
官方当前 System Card;封面日期 2026-09-22;本地 PDF/TXT:本地留存副本、current-opus-5-5-system-card.txt;230 页全文核读;当前修订时间未知
Version published
Unconfirmed
Disclosure date
Unconfirmed · Launch or later disclosure unconfirmed
Collected
Historical basis
No reliable historical version evidence available
Citation text
Claude Opus 5.5 · DRACO (100-task Perplexity deep-research benchmark), configuration DRACO low effort, in-house agent harness, Normalized score 72.5% (exact score publication date unconfirmed); Claude Opus 5.5 · DRACO (100-task Perplexity deep-research benchmark), configuration DRACO medium effort, in-house agent harness, Normalized score 83.9% (exact score publication date unconfirmed); Claude Opus 5.5 · DRACO (100-task Perplexity deep-research benchmark), configuration DRACO high effort, in-house agent harness, Normalized score 85% (exact score publication date unconfirmed); Claude Opus 5.5 · DRACO (100-task Perplexity deep-research benchmark), configuration DRACO xhigh effort, in-house agent harness, Normalized score 86.7% (exact score publication date unconfirmed); Claude Opus 5.5 · DRACO (100-task Perplexity deep-research benchmark), configuration DRACO max effort, in-house agent harness, Normalized score 87.4% (exact score publication date unconfirmed). Official source: Claude Opus 5.5 System Card (Attachment, 官方当前 System Card;封面日期 2026-09-22;本地 PDF/TXT:本地留存副本、current-opus-5-5-system-card.txt;230 页全文核读;当前修订时间未知), §8.11.2, pp.186-187, Figure 8.11.2.A, https://www.anthropic.com/claude-opus-5-5-system-card#page=187. Benchlog https://benchlog.lekshome.cc/evidence/claude-opus-5-5/ev-op55-card-draco/?lang=en
Related scores and conditions
Claude Opus 5.5 · DRACO scores and configurations
Metric and score
Configuration and evaluation conditions
Evidence relation
Normalized score72.5%Reported 72.5% · Unit %
Published Unconfirmed
DRACO low effort, in-house agent harness
Settings
Explicit low effort; web search, web fetch, programmatic tool calling, and code execution; 980k-token task budget.
Unknowns
Model run date and direct API reproduction are not stated.
Tools
Tools allowed
Reasoning budget
low
Framework
not recorded
Protocol
DRACO; low effort; web search/fetch, programmatic tools, code execution; 980k-token budget; own agent harness; five judge grading runs per response, mean; Claude Opus 4.6 grader.
Conditions
Judge choice can shift absolute scores 10-25 points; not directly comparable to paper headline scores. Cost is shown only as a chart axis and is not added as an inferred exact value. Source-wide conditions: 100 tasks; web search, web fetch, programmatic tool calling, code execution; five judge grading runs per response, mean reported; Claude Opus 4.6 grader; not directly comparable to paper headline scores.
Explicit medium effort; web search, web fetch, programmatic tool calling, and code execution; 980k-token task budget.
Unknowns
Model run date and direct API reproduction are not stated.
Tools
Tools allowed
Reasoning budget
medium
Framework
not recorded
Protocol
DRACO; medium effort; web search/fetch, programmatic tools, code execution; 980k-token budget; own agent harness; five judge grading runs per response, mean; Claude Opus 4.6 grader.
Conditions
Judge choice can shift absolute scores 10-25 points; not directly comparable to paper headline scores. Cost is not inferred from the chart. Source-wide conditions: 100 tasks; web search, web fetch, programmatic tool calling, code execution; five judge grading runs per response, mean reported; Claude Opus 4.6 grader; not directly comparable to paper headline scores.
Explicit high effort; web search, web fetch, programmatic tool calling, and code execution; 980k-token task budget.
Unknowns
Model run date and direct API reproduction are not stated.
Tools
Tools allowed
Reasoning budget
high
Framework
not recorded
Protocol
DRACO; high effort; web search/fetch, programmatic tools, code execution; 980k-token budget; own agent harness; five judge grading runs per response, mean; Claude Opus 4.6 grader.
Conditions
Judge choice can shift absolute scores 10-25 points; not directly comparable to paper headline scores. Cost is not inferred from the chart. Source-wide conditions: 100 tasks; web search, web fetch, programmatic tool calling, code execution; five judge grading runs per response, mean reported; Claude Opus 4.6 grader; not directly comparable to paper headline scores.
Explicit xhigh effort; web search, web fetch, programmatic tool calling, and code execution; 980k-token task budget.
Unknowns
Model run date and direct API reproduction are not stated.
Tools
Tools allowed
Reasoning budget
xhigh
Framework
not recorded
Protocol
DRACO; xhigh effort; web search/fetch, programmatic tools, code execution; 980k-token budget; own agent harness; five judge grading runs per response, mean; Claude Opus 4.6 grader.
Conditions
Judge choice can shift absolute scores 10-25 points; not directly comparable to paper headline scores. Cost is not inferred from the chart. Source-wide conditions: 100 tasks; web search, web fetch, programmatic tool calling, code execution; five judge grading runs per response, mean reported; Claude Opus 4.6 grader; not directly comparable to paper headline scores.
Explicit max effort; web search, web fetch, programmatic tool calling, and code execution; 980k-token task budget.
Unknowns
Model run date and direct API reproduction are not stated.
Tools
Tools allowed
Reasoning budget
max
Framework
not recorded
Protocol
DRACO; max effort; web search/fetch, programmatic tools, code execution; 980k-token budget; own agent harness; five judge grading runs per response, mean; Claude Opus 4.6 grader.
Conditions
Judge choice can shift absolute scores 10-25 points; not directly comparable to paper headline scores. Cost is not inferred from the chart. Source-wide conditions: 100 tasks; web search, web fetch, programmatic tool calling, code execution; five judge grading runs per response, mean reported; Claude Opus 4.6 grader; not directly comparable to paper headline scores.