Claude Opus 4.7 · 官方评测披露

Claude Opus 4.7 当前收录 35 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

Opus 4.7 System Card 当前保存版本 PDF pp.1-232 已逐页核读;全部图表/成绩表页已渲染并视觉核对。首发日字节尚未建立,因此 launch 状态保持未完成。 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 4.7 · 主文 · 官方发布页;本地保存 HTML/TXT 与主表 PNG:本地留存副本、本地留存副本、本地留存副本;原文发布日期 2026-04-16,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10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 4.7 System Card · 附件 · Current saved PDF/TXT: 本地留存副本 and current-opus-4-7-system-card.txt; 232 PDF pages; metadata created/modified 2026-04-16 22:11:48 China Standard Time. Full current-card review completed 2026-11-16; this file is not established as launch-day bytes.
    公布 未确认 · 采集 · 已核查 PDF 232页文本逐页全读;74页含图表或成绩表内容以160 DPI单页渲染查看;当前卡常规成绩已录入,局限与差异见独立审查报告。
  • Introducing Claude Opus 4.8 (后续成绩修订来源) · 主文 · 后续发布页;本地保存 HTML/TXT 与主表 PNG:本地留存副本、本地留存副本、本地留存副本;原文发布日期 2026-05-28,当前修订时间未知
    公布 未确认 · 采集 · 已核查 用于记录 Opus 4.7 OSWorld 更新及 Finance Agent v2 后续比较值。

配置

  • 发布表报告值 · System Card §8.1 standard configuration: adaptive thinking at max effort; default temperature/top_p; averaged over 5 trials. Context sizes are evaluation-dependent and do not exceed 1M tokens; retain all section-specific exceptions on their results. max
    未知项 Per-benchmark tools, context/token cap, resolution, trial count, harness and other exceptions are recorded on each result when stated.
  • HLE 无工具 · HLE reasoning-only, no tools; thinking auto; max reasoning effort; total tokens across contexts capped at 1M; no context compaction; Claude Opus 4.6 grader; standard five-trial average applies to the headline result. max
    未知项 The card does not state a separate headline trial count; the §8.1 standard average of five trials applies.
  • HLE 有工具 · HLE with web search, web fetch, programmatic tool calling and code execution; thinking auto; max reasoning effort; total tokens across contexts capped at 1M; no context compaction; Claude Opus 4.6 grader; known HLE sources blocklisted and transcripts reviewed. max
    未知项 The card does not state a separate headline trial count; the §8.1 standard average of five trials applies. Single-run effort-chart points use distinct configs.
  • OSWorld-Verified 更新评测 · Opus 4.8 发布说明称修改评测方式以更贴近真实使用;Opus 4.7 更新值在同文图表与脚注不一致。
    未知项 更新后的完整 harness 与测试集配置未从系统卡全文核读。
  • Finance Agent v2 后续表列值 · 由 Opus 4.8 发布表回溯列出 Opus 4.7;Finance Agent v2。
    未知项 Opus 4.7 对应的 effort/试验数未在该表说明。
  • CharXiv Reasoning,无工具 · Adaptive thinking, max effort, default sampling; 1,000 validation questions; 5-run average; current maximum image resolution (2576px single dimension, 3.75MP total). max
  • CharXiv Reasoning,Python工具 · Adaptive thinking, max effort, default sampling; Python tools; 1,000 validation questions; 5-run average; current maximum image resolution (2576px single dimension, 3.75MP total). max
  • OpenAI MRCR v2,256K bin · 8 needles; Mean Match Ratio; prompt token bin (128K, 256K]; §8.1 standard configuration unless noted.
  • OpenAI MRCR v2,1M bin · 8 needles; Mean Match Ratio; prompt token bin (524K, 1,024K]; §8.1 standard configuration unless noted.
  • USAMO 2026,medium effort · Batch API; medium effort; 300K token limit; 10 attempts per problem; six proof problems; final score is the minimum from three model judges (Gemini 3.1 Pro, Claude Opus 4.6, Claude Mythos Preview). medium
  • HLE工具评测,low effort单次 · With tools; thinking auto; low effort; one run; total token cap 1M; no context compaction. low
  • HLE工具评测,medium effort单次 · With tools; thinking auto; medium effort; one run; total token cap 1M; no context compaction. medium
  • HLE工具评测,high effort单次 · With tools; thinking auto; high effort; one run; total token cap 1M; no context compaction. high
  • HLE工具评测,xhigh effort单次 · With tools; thinking auto; xhigh effort; one run; total token cap 1M; no context compaction. xhigh
  • BrowseComp,10M token budget · Web search, web fetch, programmatic tool calling and code execution; thinking disabled; max effort; 10M total-token limit; context compaction triggered at 200K. max
  • DeepSearchQA,single-agent · 900 prompts; web search, web fetch, programmatic tool calling; adaptive thinking, max effort; context compaction at 200K; up to 10M total tokens. max
  • DRACO,max effort · 100 tasks; web search, web fetch, programmatic tool calling and code execution; adaptive thinking, max effort; 1M token limit; context compaction at 200K; five independent Opus 4.6 grading runs per response, mean normalized score. max
  • LAB-Bench FigQA,无工具 · Adaptive thinking, max effort; no tools; 5-run average; current maximum image resolution (2576px single dimension, 3.75MP total). max
  • LAB-Bench FigQA,Python工具 · Adaptive thinking, max effort; Python tools; 5-run average; current maximum image resolution (2576px single dimension, 3.75MP total). max
  • ScreenSpot-Pro,无工具 · Adaptive thinking, max effort; no tools; 1,581 tasks, 23 applications, 3 operating systems; 5-run average; current maximum image resolution (2576px single dimension, 3.75MP total). max
  • ScreenSpot-Pro,Python工具 · Adaptive thinking, max effort; Python tools; 1,581 tasks, 23 applications, 3 operating systems; 5-run average; current maximum image resolution (2576px single dimension, 3.75MP total). max
  • OfficeQA / OfficeQA Pro,exact match · System Card §8.1 standard configuration; exact-match grading; 0% allowable relative error.
  • MCP-Atlas扩展预算,max effort · Scale AI refreshed harness; adaptive thinking, max effort; 256 turns and 100 tool calls. max
  • MCP-Atlas扩展预算,high effort · Scale AI refreshed harness; adaptive thinking, high effort; 256 turns and 100 tool calls. high
  • Vending-Bench 2,max effort · One simulated business year; initial balance USD 500; final bank balance; Vending-Bench context manager used, Claude context editing disabled.
  • Vending-Bench 2,high effort · One simulated business year; initial balance USD 500; final bank balance; Vending-Bench context manager used, Claude context editing disabled.
  • GDPval-AA独立Elo评测 · Artificial Analysis independent evaluation; 220 GDPval tasks across 44 occupations and 9 industries; agent has shell and web browsing; blind pairwise output comparisons.
  • ARC-AGI-1,max effort · ARC Prize private validation set; max effort; standard sampling; score is separate from the High-effort run. max
  • ARC-AGI-1,high effort · ARC Prize private validation set; high effort; score is separate from the Max-effort summary-table value. high
  • ARC-AGI-2,max effort · ARC Prize private validation set; max thinking/effort; standard sampling. max
  • HLE工具评测,max effort单次 · With tools; thinking auto; max effort; one run; total token cap 1M; no context compaction. max
  • LAB-Bench FigQA,无工具,旧分辨率 · Adaptive thinking, max effort; no tools; 5-run average; old image limit (1568px single dimension, 1.15MP total). max
  • LAB-Bench FigQA,Python工具,旧分辨率 · Adaptive thinking, max effort; Python tools; 5-run average; old image limit (1568px single dimension, 1.15MP total). max
  • ScreenSpot-Pro,无工具,旧分辨率 · Adaptive thinking, max effort; no tools; 1,581 tasks, 23 applications, 3 operating systems; 5-run average; old image limit (1568px single dimension, 1.15MP total). max
  • ScreenSpot-Pro,Python工具,旧分辨率 · Adaptive thinking, max effort; Python tools; 1,581 tasks, 23 applications, 3 operating systems; 5-run average; old image limit (1568px single dimension, 1.15MP total). max
  • GraphWalks 256K–1M完整范围(公共 API 不可复现) · System Card §8.7.1 full-range run; 256K–1M context; 5-trial average; §8.1 standard configuration applies. Half the problems exceed the public API 1M-token limit; this config is intentionally left outside effort display modes.
    未知项 Section 8.7.1 does not restate API route or per-run token accounting.
  • OfficeQA Pro current announcement update · Claude Opus 4.7; announcement figure reported setting; per-result effort and evaluator not specified.
    未知项 题集、harness、effort和trials未在该图注明,不能与系统卡的其他运行设置强并。
  • Structural Biology announcement reported setting · Claude Opus 4.7; announcement figure reported setting; per-result effort and evaluator not specified.
    未知项 题集、harness、effort和trials未在该图注明,不能与系统卡的其他运行设置强并。

直接前代与后续发布

直接前代 Claude Opus 4.6。 官方关系依据

Opus 4.7 is a direct upgrade to Opus 4.6.
17 项披露延续 · 18 项新增 · 18 项本次未见 · 0 项待核查
Claude Opus 4.6 与 Claude Opus 4.7 的披露变化
测试已核查状态Claude Opus 4.6Claude Opus 4.7
SWE-bench Pro · Pro新增—本次记录
SWE-bench Verified · Verified subset (500 problems)延续

已记录的评测条件有变化

前代记录本次记录
Terminal-Bench · 2.0;代理框架按协议区分延续

已记录的评测条件有变化

前代记录本次记录
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
BrowseComp · 1,266 题;单代理与多代理按协议区分延续

已记录的评测条件有变化

前代记录本次记录
MCP-Atlas · Scale AI 公共排行榜;2026-04 刷新 harness新增—本次记录
OSWorld-Verified · Verified benchmark; evaluation protocol noted per source延续

已记录的评测条件有变化

前代记录本次记录
Finance Agent · Vals AI SEC-filing research benchmark新增—本次记录
GPQA Diamond · 198-question Diamond subset延续

已记录的评测条件有变化

前代记录本次记录
CharXiv Reasoning · Validation split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
MMMLU · 57 subjects and 14 non-English languages延续

已记录的评测条件有变化

前代记录本次记录
Finance Agent · v2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新增—本次记录
GraphWalks · 200 problems, 256k to 1M context, BFS F1新增—本次记录
OpenAI MRCR · v2, 256K context, 8 needles新增—本次记录
OpenAI MRCR · v2, 1M context, 8 needles新增—本次记录
USA Mathematical Olympiad · 2026新增—本次记录
DeepSearchQA · 900 prompts;单代理与多代理按协议区分延续

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
DRACO · 100-task Perplexity deep-research benchmark新增—本次记录
LAB-Bench FigQA · FutureHouse FigQA; prompting and tools are recorded per result延续

已记录的评测条件有变化

前代记录本次记录
ScreenSpot-Pro · 1,581 expert-annotated tasks across 23 professional applications and 3 operating systems新增—本次记录
OfficeQA · Full set新增—本次记录
Vending-Bench · 2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
GDPval-AA · Artificial Analysis leaderboard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
ARC-AGI-1 · ARC Prize Foundation private dataset新增—本次记录
ARC-AGI-2 · ARC Prize Foundation private dataset延续

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
INCLUDE · 44 languages新增—本次记录
GraphWalks · parents;256K–1M 上下文;题数未注明延续

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
GraphWalks · 256K-context subset of the 256K-1M evaluation; item count unspecified; BFS F1延续

已记录的评测条件有变化

前代记录本次记录
GraphWalks · 256K-context subset of the 256K-1M evaluation; item count unspecified; parents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新增—本次记录
Structural Biology · Opus 4.7 announcement figure; dataset, version and question format unspecified新增—本次记录
τ²-bench · Retail section本次未见前代记录—
τ²-bench · Telecom section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—
Finance Agent · Opus 4.6 来源未注明版本;发布于 Vals AI v2 (2026-05) 之前本次未见前代记录—
ARC-AGI-2 · Verified;题集组成未在发布表中说明本次未见前代记录—
MMMU-Pro · Standard and Vision formats本次未见前代记录—
OpenRCA · 335 enterprise software failure cases本次未见前代记录—
AIME · 2025 edition本次未见前代记录—
Real-World Finance · Anthropic internal evaluation; about 50 analyst-workflow tasks; version unspecified本次未见前代记录—
OpenAI MRCR v2 · v2 fix dated 2025-12-05; 8 needles; prompts in (128K, 256K] tokens; Mean Match Ratio本次未见前代记录—
OpenAI MRCR v2 · v2 fix dated 2025-12-05; 8 needles; prompts in (524K, 1024K] tokens; full variant may exceed the public API context limit本次未见前代记录—
GraphWalks · BFS;256K–1M 上下文;题数未注明本次未见前代记录—
WebArena · Official WebArena grader, modified fuzzy_match subgrader本次未见前代记录—
BioPipelineBench · Anthropic internal unpublished computational-biology evaluation; version and task count unspecified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—
Structural Biology · Anthropic internal unpublished life-science evaluation; format recorded per result本次未见前代记录—
Organic Chemistry · Anthropic internal unpublished life-science evaluation; version unspecified本次未见前代记录—
Phylogenetics · Anthropic internal unpublished life-science evaluation; version unspecified本次未见前代记录—

测试记录 · 35