Claude Opus 5.5 · 官方评测披露

Claude Opus 5.5 当前收录 41 项测试的官方记录。已完成列明材料的核查。

模型发布 · Claude Opus

官方材料与核查范围

Current Opus 5.5 System Card pp.1-230 and the official release table are complete; the Haiku 5.5 comparison card pp.1-144 is complete from the combined Singer/Boyle review; the later Fable 5.2 article and all nine charts were fully reviewed by the parent. The Fable comparison is later evidence and does not change launch availability. 核对时间 。

当前资料不足以还原首发。 没有发布当日历史快照。

  • Introducing Claude Opus 5.5 · 主文 · 官方发布页;本地保存 HTML/TXT:本地留存副本、本地留存副本;主表为 HTML;原文发布日期 2026-09-22,当前修订时间未知
    公布 未确认 · 采集 · 已核查 已补完当前发布主文全文、全部引言与脚注,实际查看全部13个figure及正文/品牌图像。核读与图片清单见本地留存副本。未估读无数值曲线,客户私有评测或缺方法的引述不自动作为普通能力成绩。
  • 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
    公布 未确认 · 采集 · 已核查 当前保存版 230 页全文核读;230 个 TXT 页段连续且无替换字符;63 张双页接触表覆盖 126 个含图页,全部逐张实际查看,并放大查看原 PDF pp.187-188。
  • Claude Haiku 5.5 System Card (Opus 5.5 OSWorld offline comparison) · 附件 · Official current System Card, cover dated 2026-10-07; current PDF/TXT saved at 本地留存副本 and .txt; full pp.1-144 read across Singer/Boyle reviews, with 63 safety figure pages visually checked; current revision date unknown.
    公布 未确认 · 采集 · 已核查 Singer/Boyle combined current-card review covers all pp.1-144; 144 continuous TXT page segments; 63 safety figure pages visually checked; pp.111-139 capability section was previously read by Boyle. This completes the auxiliary source review; it does not establish launch-day bytes.
  • Introducing Claude Fable 5.2 and Claude Mythos 5.2 · 主文 · Official article published 2026-11-09 and edited 2026-11-11; current HTML/TXT saved at 本地留存副本 and .txt; later Opus 5.5 comparison, not launch-day evidence.
    公布 · 采集 · 已核查 Parent completed the full official Fable 5.2 article and viewed all nine charts; RoboDojo chart 8 was visually verified for the Opus 5.5 39.4% Max-effort bar and internal-run/highest-effort footnote. This is later-comparison evidence only.

配置

  • Adaptive Max · 官方主表标准配置为 adaptive thinking/max effort、默认 temperature/top_p;具体工具、safeguards、fallback 与运行数按各评测章节记录,不将其一概视为生产 safeguards enabled。 max
    未知项 表格没有为每个 benchmark 单独列出采样设置和运行数;章节明确列出的差异优先。
  • Terminal-Bench xhigh · Terminal-Bench 4.0 使用 xhigh effort;官方说明这是 Opus 5.5 的最高分设置。 xhigh
    未知项 每题运行数未在该主表脚注具体给出。
  • HLE 有工具 · Humanity's Last Exam; with tools; adaptive thinking/max effort per table note。 max
    未知项 具体题集和 trial count 未列。
  • System Card OSWorld 2.0 更新协议 · 108 个 long-horizon desktop tasks;五次独立运行;1080p、最多 500 steps、max effort;更新 task/assets 与 harness。 max
    未知项 Card 当前修订日期未公布;与主文 OSWorld 2.1 标签对应关系不明确。
  • OSWorld 2.1 offline subset · Haiku 5.5 System Card 报告 82/108 offline tasks、no internet、五次尝试;max effort。 max
    未知项 该卡当前修订日期未公布。
  • System Card max effort (per-result benchmark harnesses) · System Card §8.1 default: adaptive thinking/max effort and default sampling settings unless a section says otherwise. Benchmark tools, evaluator, safeguards, fallback, and repetitions remain per-result conditions; this config represents the stated model effort, not harness equivalence. max
    未知项 Exact run time is not disclosed.、Per-benchmark tools, safeguards, and repetition counts are specified only where the source states them.
  • FrontierCode v1.1 Main medium effort · Explicit medium effort; Cognition evaluation in Claude Code. Mean@5. medium
    未知项 No API-equivalent harness or run date is stated.
  • FrontierCode v1.1 Extended medium effort · Explicit medium effort; Cognition evaluation in Claude Code. Mean@5. medium
    未知项 No API-equivalent harness or run date is stated.
  • FrontierCode v1.1 Extended max effort · Explicit max effort; Cognition evaluation in Claude Code. Mean@5. max
    未知项 No API-equivalent harness or run date is stated.
  • CursorBench 4.0 medium effort · Explicit medium effort in Cursor's production agent harness; Cursor measured the score. medium
    未知项 No API-equivalent harness or run date is stated.
  • CursorBench 4.0 high effort · Explicit high effort in Cursor's production agent harness; Cursor measured the score. high
    未知项 No API-equivalent harness or run date is stated.
  • CursorBench 4.0 xhigh effort · Explicit xhigh effort in Cursor's production agent harness; Cursor measured the score. xhigh
    未知项 No API-equivalent harness or run date is stated.
  • HLE no tools, low effort · Reasoning-only; explicit low effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader. low
    未知项 The figure does not state the number of model trials.
  • HLE no tools, medium effort · Reasoning-only; explicit medium effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader. medium
    未知项 The figure does not state the number of model trials.
  • HLE no tools, high effort · Reasoning-only; explicit high effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader. high
    未知项 The figure does not state the number of model trials.
  • HLE no tools, xhigh effort · Reasoning-only; explicit xhigh effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader. xhigh
    未知项 The figure does not state the number of model trials.
  • HLE no tools, max effort · Reasoning-only; explicit max effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader. max
    未知项 The figure does not state the number of model trials.
  • HLE with tools, low effort · Web search, web fetch, programmatic tool calling, and code execution; explicit low effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader. low
    未知项 The figure does not state the number of model trials.
  • HLE with tools, medium effort · Web search, web fetch, programmatic tool calling, and code execution; explicit medium effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader. medium
    未知项 The figure does not state the number of model trials.
  • HLE with tools, high effort · Web search, web fetch, programmatic tool calling, and code execution; explicit high effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader. high
    未知项 The figure does not state the number of model trials.
  • HLE with tools, xhigh effort · Web search, web fetch, programmatic tool calling, and code execution; explicit xhigh effort; adaptive thinking; 1M total-token cap; no context compaction; Claude Opus 4.6 grader. xhigh
    未知项 The figure does not state the number of model trials.
  • DRACO low effort, in-house agent harness · Explicit low effort; web search, web fetch, programmatic tool calling, and code execution; 980k-token task budget. low
    未知项 Model run date and direct API reproduction are not stated.
  • DRACO medium effort, in-house agent harness · Explicit medium effort; web search, web fetch, programmatic tool calling, and code execution; 980k-token task budget. medium
    未知项 Model run date and direct API reproduction are not stated.
  • DRACO high effort, in-house agent harness · Explicit high effort; web search, web fetch, programmatic tool calling, and code execution; 980k-token task budget. high
    未知项 Model run date and direct API reproduction are not stated.
  • DRACO xhigh effort, in-house agent harness · Explicit xhigh effort; web search, web fetch, programmatic tool calling, and code execution; 980k-token task budget. xhigh
    未知项 Model run date and direct API reproduction are not stated.
  • DRACO max effort, in-house agent harness · Explicit max effort; web search, web fetch, programmatic tool calling, and code execution; 980k-token task budget. max
    未知项 Model run date and direct API reproduction are not stated.
  • WANDR low effort, frozen-index harness · Explicit low effort; web search and web fetch limited to a frozen index; programmatic tool calling and code execution; 980k-token task budget. low
    未知项 Model run date and direct API reproduction are not stated.
  • WANDR medium effort, frozen-index harness · Explicit medium effort; web search and web fetch limited to a frozen index; programmatic tool calling and code execution; 980k-token task budget. medium
    未知项 Model run date and direct API reproduction are not stated.
  • WANDR high effort, frozen-index harness · Explicit high effort; web search and web fetch limited to a frozen index; programmatic tool calling and code execution; 980k-token task budget. high
    未知项 Model run date and direct API reproduction are not stated.
  • WANDR xhigh effort, frozen-index harness · Explicit xhigh effort; web search and web fetch limited to a frozen index; programmatic tool calling and code execution; 980k-token task budget. xhigh
    未知项 Model run date and direct API reproduction are not stated.
  • WANDR max effort, frozen-index harness · Explicit max effort; web search and web fetch limited to a frozen index; programmatic tool calling and code execution; 980k-token task budget. max
    未知项 Model run date and direct API reproduction are not stated.
  • Chartography max effort, no tools · Adaptive thinking/max effort; no tools; five runs; Gemini 3.5 Flash grader. max
  • Chartography max effort, with tools · Adaptive thinking/max effort; container, standard libraries, and image-cropping tool; five runs; Gemini 3.5 Flash grader. max
  • BenchCAD Vision2Code max effort, no tools · Adaptive thinking/max effort; corrected 256 x 256 px per view implementation; random 1,000-file subset; five runs. max
    未知项 Task-level variance is not separately reported in the card text.
  • BenchCAD Vision2Code max effort, with tools · Adaptive thinking/max effort; container, standard libraries, and image-cropping tool; corrected 256 x 256 px per view implementation; random 1,000-file subset; five runs. max
    未知项 Task-level variance is not separately reported in the card text.
  • GDPval-AA v2.1 xhigh effort · Explicit xhigh effort; Artificial Analysis independent harness; 220 tasks across 44 occupations. xhigh
    未知项 Run date is not stated.
  • AA-Briefcase v1.1 high effort · Explicit high effort; Artificial Analysis independent harness. high
    未知项 Run date is not stated.
  • AA-Briefcase v1.1 xhigh effort · Explicit xhigh effort; Artificial Analysis independent harness. xhigh
    未知项 Run date is not stated.
  • AA-Briefcase v1.1 max effort · Explicit max effort; Artificial Analysis independent harness. max
    未知项 Run date is not stated.
  • Toolathlon-Verified max effort with safeguards · Adaptive thinking/max effort; production safety classifiers and sandbox-escape monitor enabled; three trials per task; all stopped trials counted as failures. max
  • HealthBench max effort with safeguards · Adaptive thinking/max effort; safety classifiers enabled and refusal fallback to Opus 5; no tools or customized system prompts; five trials. max
  • HealthBench Professional max effort with safeguards · Adaptive thinking/max effort; safety classifiers enabled and refusal fallback to Opus 5; no tools or customized system prompts; five trials. max
  • RoboDojo Sim internal run, max effort · Official Fable 5.2 article chart labels Opus 5.5 Max effort; footnote says RoboDojo Sim used internally run models at their highest-available effort. max
    未知项 RoboDojo version, task count, prompt/harness, trial count, and run date are not stated.

直接前代与后续发布

直接前代 Claude Opus 5。 官方关系依据

Opus 5.5 is a major step up from Opus 5.
25 项披露延续 · 16 项新增 · 21 项本次未见 · 0 项待核查
Claude Opus 5 与 Claude Opus 5.5 的披露变化
测试已核查状态Claude Opus 5Claude Opus 5.5
Terminal-Bench · 4.0新增—本次记录
FrontierCode · v1.1 Main set延续

已记录的评测条件有变化

前代记录本次记录
CursorBench · 4.0新增—本次记录
GDPval-AA · v2.1新增—本次记录
AutomationBench · Zapier 私有 early-access set;版本未注明延续

已记录的评测条件有变化

前代记录本次记录
Humanity's Last Exam · 2,500-question benchmark延续

已记录的评测条件有变化

前代记录本次记录
Terminal-Bench-Science · 0.1新增—本次记录
OSWorld · 2.1 (September 10, 2026 task/assets release)新增—本次记录
Chartography · 100 tasks延续

来源位置 attachment → attachment、main

已记录的评测条件有变化

前代记录本次记录
OSWorld · System Card 所称 2.0;2026-09-10 任务与资源版本、108 题、更新 harness延续

已记录的评测条件有变化

前代记录本次记录
OSWorld · 2.1, 82-task offline no-internet subset, partial score新增—本次记录
SWE-bench Pro · Pro延续

已记录的评测条件有变化

前代记录本次记录
DeepSWE · v1.1延续

来源位置 attachment、main → attachment

已记录的评测条件有变化

前代记录本次记录
FrontierSWE · v2新增—本次记录
DRACO · 100-task Perplexity deep-research benchmark延续

已记录的评测条件有变化

前代记录本次记录
WANDR · 500-task wide research benchmark新增—本次记录
ProgramBench · 166 个保留的 golden tasks;架构与时限按协议区分延续

已记录的评测条件有变化

前代记录本次记录
BenchCAD Vision2Code · 1,000-file random subset of 17,900 files延续

已记录的评测条件有变化

前代记录本次记录
OfficeQA · Full set延续

已记录的评测条件有变化

前代记录本次记录
Legal Agent Benchmark · Opus 5 官方发布表所列 held-out set延续

来源位置 attachment、main → attachment

已记录的评测条件有变化

前代记录本次记录
AA-Briefcase · v1.1新增—本次记录
Toolathlon-Verified · June 2026 finalized release新增—本次记录
HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations延续

已记录的评测条件有变化

前代记录本次记录
HealthBench Professional · 525 physician-authored conversations延续

来源位置 attachment、main → attachment

已记录的评测条件有变化

前代记录本次记录
Global MMLU · 42 languages including English · overall average延续

已记录的评测条件有变化

前代记录本次记录
Multi-task Indic Language Understanding Benchmark · 10 Indic languages plus English · overall average延续

已记录的评测条件有变化

前代记录本次记录
SpatialBench Verified · Verified延续

已记录的评测条件有变化

前代记录本次记录
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延续

已记录的评测条件有变化

前代记录本次记录
De novo protein binder design · System Card evaluation; 15 targets; 30 miniprotein binders per target; Cas9 excluded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延续

已记录的评测条件有变化

前代记录本次记录
RoboDojo Sim · 版本/任务配置未注明新增—本次记录
SWE-bench Multilingual · 300 problems, 9 programming languages延续

已记录的评测条件有变化

前代记录本次记录
SWE-bench Multimodal · Multimodal延续

已记录的评测条件有变化

前代记录本次记录
FrontierCode · v1.1 Extended set延续

已记录的评测条件有变化

前代记录本次记录
OfficeQA · Pro 133-question subset延续

已记录的评测条件有变化

前代记录本次记录
SingleCellBench · LatchBio Bioinformatics single-cell RNA sequencing analysis evaluation延续

已记录的评测条件有变化

前代记录本次记录
Protein Design Library Ranking · Prioritization of protein designs for experimental testing新增—本次记录
Protocol Understanding · V2; Benchling-created evaluation using real laboratory protocols and applied questions; task count unspecified延续

已记录的评测条件有变化

前代记录本次记录
Frontier-Bench · v0.1本次未见前代记录—
GDPval-AA · v2本次未见前代记录—
ARC-AGI-3 · 版本/子集未注明本次未见前代记录—
BrowseComp · 1,266 题;单代理与多代理按协议区分本次未见前代记录—
OSWorld · 2.0;Opus 5 发布表,未注明评测子集/协议本次未见前代记录—
BioMysteryBench · 官方发布表列 Hard 与 human-solved 结果;细分版本未注明本次未见前代记录—
SWE-bench Verified · Verified subset (500 problems)本次未见前代记录—
International Mathematical Olympiad · 2026 contest; six proof problems本次未见前代记录—
RiemannBench · 25 private research-mathematics problems; corrected reference answers and grading; four attempts per problem本次未见前代记录—
ArXivMath · June 2026 release (49 problems)本次未见前代记录—
DeepSearchQA · 900 prompts;单代理与多代理按协议区分本次未见前代记录—
GDP.pdf · 100 professional PDFs and prompts本次未见前代记录—
MCP-Atlas · 所选材料未明确子集本次未见前代记录—
Legal Agent Benchmark · Public full set本次未见前代记录—
AA-Briefcase · Not stated in selected source本次未见前代记录—
ARC-AGI-1 · ARC Prize Foundation semi-private 2026 evaluation set本次未见前代记录—
ARC-AGI-2 · ARC Prize Foundation semi-private 2026 evaluation set本次未见前代记录—
ARC-AGI-3 · ARC Prize Foundation semi-private 2026 evaluation set本次未见前代记录—
INCLUDE · 44 languages本次未见前代记录—
ProteinGym Hard · System-card Hard subset; exact task count not stated本次未见前代记录—
Organic Chemistry · V2; system-card evaluation with expert-recommended harder problems本次未见前代记录—

测试记录 · 41