Claude Haiku 5.5 · 官方评测披露

Claude Haiku 5.5 当前收录 33 项测试的官方记录。已完成列明材料的核查。

模型发布 · Claude Haiku

官方材料与核查范围

当前官方发布页及 144 页 System Card 均已完成核读。System Card pp.1-144 全文已逐页核读,全部63个含图/向量图表页已实际视觉检查;未读页或未查看图表页为零。 核对时间 。

当前资料不足以还原首发。 本地 CDX 查询记录中发布当日主文与系统卡均无可回放 HTTP 200 快照;本次对 web.archive.org 的 CDX/回放重试分别遇到 TLS EOF 与连接超时。当前主文和系统卡不作为首发版本证据。

  • Introducing Claude Haiku 5.5 · 主文 · 官方发布页;本次保存 HTML/TXT:本地留存副本、current-haiku-5-5-main.txt;发布日期 2026-10-07,当前修订时间未知
    公布 · 采集 · 已核查 本次保存的官方 HTML/TXT 已核读;主表、effort 图表说明、价格、安全摘要、availability 与脚注均已读取。
  • Claude Haiku 5.5 System Card · 附件 · 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。
    公布 · 采集 · 已核查 PDF pp.1-144 已全文核读。能力章节 pp.111-139 已单独核读;其余 pp.1-110、140-144 已补核。全部63个含图/向量图表页已实际视觉检查:本轮新增53页,合并此前10页;精确清单见 本地留存副本。未从无标签图形估值。

配置

  • 发布表 Max effort · Anthropic 主文表格;unless stated otherwise, adaptive thinking and max effort。 max
    未知项 逐 benchmark sampling settings 和 trial counts 未在主表完整列出。
  • OSWorld 2.1 offline subset · OSWorld 2.1 official offline subset: 82 of 108 tasks, no internet, 1080p, max 500 action steps; partial score and strict pass rate are pass@1 averaged over five independent attempts per task at max effort. max
    未知项 System Card 当前修订日期未给出。
  • System Card 标准 max effort · Unless otherwise noted: adaptive thinking at max effort, default temperature/top_p, five-trial average; context window is evaluation-dependent and no more than 1M tokens. max
    未知项 每个 benchmark 的具体 context window 未逐项列明。
  • HLE with tools / low · Thinking auto; low effort; web search, web fetch, programmatic tool calling, and code execution; 980k-token task budget; no context compaction; Opus 4.6 grader. low
    未知项 Per-effort trial count is not stated in the chart caption where applicable.
  • HLE without tools / low · Thinking auto; low effort; no tools; 980k-token task budget; no context compaction. low
    未知项 Per-effort trial count is not stated in the chart caption where applicable.
  • HLE with tools / medium · Thinking auto; medium effort; web search, web fetch, programmatic tool calling, and code execution; 980k-token task budget; no context compaction; Opus 4.6 grader. medium
    未知项 Per-effort trial count is not stated in the chart caption where applicable.
  • HLE without tools / medium · Thinking auto; medium effort; no tools; 980k-token task budget; no context compaction. medium
    未知项 Per-effort trial count is not stated in the chart caption where applicable.
  • HLE with tools / high · Thinking auto; high effort; web search, web fetch, programmatic tool calling, and code execution; 980k-token task budget; no context compaction; Opus 4.6 grader. high
    未知项 Per-effort trial count is not stated in the chart caption where applicable.
  • HLE without tools / high · Thinking auto; high effort; no tools; 980k-token task budget; no context compaction. high
    未知项 Per-effort trial count is not stated in the chart caption where applicable.
  • HLE with tools / xhigh · Thinking auto; xhigh effort; web search, web fetch, programmatic tool calling, and code execution; 980k-token task budget; no context compaction; Opus 4.6 grader. xhigh
    未知项 Per-effort trial count is not stated in the chart caption where applicable.
  • HLE without tools / xhigh · Thinking auto; xhigh effort; no tools; 980k-token task budget; no context compaction. xhigh
    未知项 Per-effort trial count is not stated in the chart caption where applicable.
  • HLE with tools / max · Thinking auto; max effort; web search, web fetch, programmatic tool calling, and code execution; 980k-token task budget; no context compaction; Opus 4.6 grader. max
    未知项 Per-effort trial count is not stated in the chart caption where applicable.
  • HLE without tools / max · Thinking auto; max effort; no tools; 980k-token task budget; no context compaction. max
    未知项 Per-effort trial count is not stated in the chart caption where applicable.
  • System Card effort / low · Adaptive thinking; low effort, explicitly labeled in the source; benchmark-specific protocol is preserved in each result's conditions. low
    未知项 Sampling and trial counts vary by evaluation and are recorded per result.
  • System Card effort / medium · Adaptive thinking; medium effort, explicitly labeled in the source; benchmark-specific protocol is preserved in each result's conditions. medium
    未知项 Sampling and trial counts vary by evaluation and are recorded per result.
  • System Card effort / high · Adaptive thinking; high effort, explicitly labeled in the source; benchmark-specific protocol is preserved in each result's conditions. high
    未知项 Sampling and trial counts vary by evaluation and are recorded per result.
  • System Card effort / xhigh · Adaptive thinking; xhigh effort, explicitly labeled in the source; benchmark-specific protocol is preserved in each result's conditions. xhigh
    未知项 Sampling and trial counts vary by evaluation and are recorded per result.
  • System Card effort / max · Adaptive thinking; max effort, explicitly labeled in the source; benchmark-specific protocol is preserved in each result's conditions. max
    未知项 Sampling and trial counts vary by evaluation and are recorded per result.
  • Healthcare evaluations / low · Safety classifiers enabled; low effort; run counts, scoring variants, and dates are recorded per result. low
    未知项 Healthcare evaluations use benchmark-specific run counts and scoring variants.
  • Healthcare evaluations / medium · Safety classifiers enabled; medium effort; run counts, scoring variants, and dates are recorded per result. medium
    未知项 Healthcare evaluations use benchmark-specific run counts and scoring variants.
  • Healthcare evaluations / high · Safety classifiers enabled; high effort; run counts, scoring variants, and dates are recorded per result. high
    未知项 Healthcare evaluations use benchmark-specific run counts and scoring variants.
  • Healthcare evaluations / xhigh · Safety classifiers enabled; xhigh effort; run counts, scoring variants, and dates are recorded per result. xhigh
    未知项 Healthcare evaluations use benchmark-specific run counts and scoring variants.
  • Healthcare evaluations / max · Safety classifiers enabled; max effort; run counts, scoring variants, and dates are recorded per result. max
    未知项 Healthcare evaluations use benchmark-specific run counts and scoring variants.
  • GDPval-AA v2.1 medium effort · Medium effort; score run independently by Artificial Analysis. Medium
    未知项 Trial count and remaining sampling details are not stated in the System Card.
  • AA-Briefcase v1.1 medium effort · Medium effort; score run independently by Artificial Analysis. Medium
    未知项 Trial count and remaining sampling details are not stated in the System Card.
  • Life sciences capabilities · Adaptive thinking at max effort; biology safeguards disabled; five attempts per problem unless an evaluation subsection states otherwise. max
    未知项 Per-benchmark sample counts are not stated for all evaluations.

直接前代与后续发布

直接前代 Claude Haiku 4.5。 官方关系依据

We present results from a wide variety of pre-deployment evaluations, which show Claude Haiku 5.5 significantly outperforming its predecessor, Claude Haiku 4.5, across many domains.
0 项披露延续 · 33 项新增 · 10 项本次未见 · 0 项待核查
Claude Haiku 4.5 与 Claude Haiku 5.5 的披露变化
测试已核查状态Claude Haiku 4.5Claude Haiku 5.5
GDPval-AA · v2.1新增—本次记录
AA-Briefcase · v1.1新增—本次记录
OSWorld · 2.1, 82-task offline no-internet subset, partial score新增—本次记录
Humanity's Last Exam · 2,500-question benchmark新增—本次记录
Terminal-Bench · 4.0新增—本次记录
FrontierCode · v1.1 Main set新增—本次记录
Chartography · 100 tasks新增—本次记录
SWE-bench Pro · Pro新增—本次记录
SWE-bench Multilingual · 300 problems, 9 programming languages新增—本次记录
SWE-bench Multimodal · Multimodal新增—本次记录
FrontierCode · v1.1 Extended set新增—本次记录
Terminal-Bench-Science · 0.1新增—本次记录
FrontierSWE · v2新增—本次记录
ProgramBench · 166 个保留的 golden tasks;架构与时限按协议区分新增—本次记录
DRACO · 100-task Perplexity deep-research benchmark新增—本次记录
WANDR · 500-task wide research benchmark新增—本次记录
BenchCAD Vision2Code · 1,000-file random subset of 17,900 files新增—本次记录
OfficeQA · Full set新增—本次记录
OfficeQA · Pro 133-question subset新增—本次记录
HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations新增—本次记录
HealthBench Professional · 525 physician-authored conversations新增—本次记录
PhysicianBench · 100 physician tasks in an electronic health record; pass rate across attempts新增—本次记录
Global MMLU · 42 languages including English · overall average新增—本次记录
Multi-task Indic Language Understanding Benchmark · 10 Indic languages plus English · overall average新增—本次记录
SpatialBench Verified · Verified新增—本次记录
SingleCellBench · LatchBio Bioinformatics single-cell RNA sequencing analysis evaluation新增—本次记录
Morphology-to-molecule matching · Four multi-channel human-liver in-vitro cell imaging sets matched to shuffled molecule-dose perturbations新增—本次记录
Medicinal Chemistry ADME · 504 drug-discovery questions about measured ADME properties新增—本次记录
Protein Design Sequence Generation · Protein sequence generation conditioned on design constraints新增—本次记录
Protein Design Library Ranking · Prioritization of protein designs for experimental testing新增—本次记录
Biomedical Image Analysis · System-card evaluation across public biomedical-image competition tasks; included competition set unspecified新增—本次记录
Protocol Troubleshooting · System-card protocol troubleshooting variant; task count unspecified新增—本次记录
Protocol Understanding · V2; Benchling-created evaluation using real laboratory protocols and applied questions; task count unspecified新增—本次记录
SWE-bench Verified · Verified subset (500 problems)本次未见前代记录—
Terminal-Bench · 版本未注明;Claude Haiku 4.5 发布材料所列本次未见前代记录—
τ²-bench · Retail section本次未见前代记录—
tau2-bench · Airline category本次未见前代记录—
τ²-bench · Telecom section本次未见前代记录—
OSWorld-Verified · Verified benchmark; evaluation protocol noted per source本次未见前代记录—
AIME · 2025 edition本次未见前代记录—
GPQA Diamond · 198-question Diamond subset本次未见前代记录—
MMMLU · 57 subjects and 14 non-English languages本次未见前代记录—
MMMU · Validation split本次未见前代记录—

测试记录 · 33