Claude Opus 4.5 · 官方评测披露

Claude Opus 4.5 当前收录 31 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 4.5 与后续 Opus 4.6 当前发布文章均全文核读,所有 figure 和图像均实际查看;当前保存版 Opus 4.5 System Card 共 153 页,正文按页 1–153 全读,全部含表格或图表的页面均实际渲染核验。此完成状态仅适用于已保存版本,不证明其等同首发附件。 核对时间 。

当前资料不足以还原首发。 未取得独立的发布当日历史快照;发布页当前版本不能证明历史页面状态。

  • Introducing Claude Opus 4.5 · 主文 · 官方发布页;原文发布日期 2025-11-24。保留的发布材料:本地留存副本、本地留存副本、本地留存副本。当前 HTML/TXT:本地留存副本 与 current-opus-4-5-main.txt;HTML last-modified 2026-10-14 23:48:16 UTC。2026-11-16 已全文核读并检查全部 12 个 figure、21 个非 figure 图像;截图清单见 本地留存副本。当前页不证明首发附件内容。
    公布 未确认 · 采集 · 已核查 current-opus-4-5-main.txt 从首行至末行全文核读;current-opus-4-5-main.html 逐段核对;全部 12 个 figure、21 个非 figure 图像及复用的主表比较 PNG 均实际视觉检查。未读段落/图像为零;截图清单见 本地留存副本。Current HTML last-modified 2026-10-14 23:48:16 UTC;首发历史状态未知。
  • Introducing Claude Opus 4.6 · 主文 · 后续对照来源;原文发布日期 2026-02-05。当前 HTML/TXT:本地留存副本 与 current-opus-4-6-main.txt;HTML last-modified 2026-10-14 23:55:44 UTC。2026-11-16 已全文核读并检查全部 14 个 figure、21 个非 figure 图像;截图清单见 本地留存副本。仅按该后续来源日期记录对照披露,不回填 Opus 4.5 首发日。
    公布 未确认 · 采集 · 已核查 作为 Opus 4.5 的后续对照来源,current-opus-4-6-main.txt 从首行至末行全文核读;current-opus-4-6-main.html 逐段核对;全部 14 个 figure、21 个非 figure 图像及复用的主表比较 PNG 均实际视觉检查。未读段落/图像为零;截图清单见 本地留存副本。发布页日期 2026-02-05,current HTML last-modified 2026-10-14 23:55:44 UTC。
  • Claude Opus 4.5 System Card · 附件 · Current saved PDF/TXT: 本地留存副本 and .txt; cover November 2025, change log 2025-11-24; current binary is not established as launch-day bytes; fully reviewed 2026-11-16.
    公布 未确认 · 采集 · 已核查 当前保存版 PDF/TXT 共 153 页,页段连续且无缺页;正文 1–153 已逐页核读,所有表格/图表页均实际查看渲染;无嵌入附件。视觉页完整清单见独立目录 source-review-patch.json。

配置

  • 默认 High · 64K thinking budget、200K context、默认 high、默认 sampling、5 次独立试验。 high · 思考预算 64K
  • SWE-bench 无 thinking budget · 发布材料明确说明 SWE-bench Verified 不使用 thinking budget。
    未知项 除发布表和脚注外未提取完整测试脚手架细节。
  • Terminal-Bench 128K thinking budget · 默认 Terminus 2;128K thinking budget。 思考预算 128K
    未知项 文章未说明是否与其它行采用相同试验次数。
  • No extended thinking · Extended thinking off; 200K context and default sampling are stated for MCP Atlas; Table 2.3.A applies its no-thinking footnote to MCP Atlas, τ² Retail and τ² Telecom.
    未知项 The system card does not state an independent effort level for these no-thinking evaluations.
  • Terminal-Bench 2.0, 64K thinking budget · Terminus-2 harness in Harbor scaffold; 64K thinking budget. 思考预算 64K
    未知项 The system card does not separately state effort or default sampling settings for this run.
  • Vending-Bench 2, High effort · High effort; 8,192 reasoning tokens per turn; benchmark-managed context, Claude context editing disabled. high · 思考预算 8,192 reasoning tokens per turn
    未知项 Run count and sampling settings are not stated.
  • ARC Prize Foundation private validation, 64K thinking · 64K thinking tokens; ARC Prize Foundation private validation dataset. 思考预算 64K
    未知项 Run count and evaluation harness details are not stated in the card.
  • HLE reasoning-only, no tools · Reasoning-only evaluation without tools; 2,500-question HLE benchmark.
    未知项 Reasoning budget, trial count and other sampling settings are not stated.
  • HLE tools-only, no reasoning · Web search, web fetch and code execution; no reasoning.
    未知项 Trial count and detailed tool budgets are not stated.
  • FinanceAgent external Vals AI evaluation · 64K thinking budget; 200K context; external Vals AI analysis; eight-trial average. 思考预算 64K
    未知项 The system card does not state a named FinanceAgent version or effort label.
  • SpreadsheetBench, no extended thinking · Full 912-problem set; no extended thinking; 200K context; custom bash/string/Python harness; five-trial average.
    未知项 No further prompt or sampling details are stated.
  • AIME 2025, no tools · 64K thinking budget; interleaved scratchpads; 200K context; default High effort and sampling; five-trial average; no tools. high · 思考预算 64K
    未知项 The card warns contamination may have inflated the result.
  • AIME 2025, Python tools · 64K thinking budget; interleaved scratchpads; 200K context; default High effort and sampling; Python tools; five-trial average. high · 思考预算 64K
    未知项 Exact Python harness details are not stated; the card warns contamination may have inflated the result.
  • LAB-Bench FigQA, no tools · Extended thinking off; no tools; zero-shot prompting.
    未知项 Trial count is not stated.
  • LAB-Bench FigQA, image cropping tool · Simple image cropping tool; 32,768 reasoning-token budget; zero-shot prompting. 思考预算 32,768
    未知项 Trial count is not stated.
  • WebArena single-policy evaluation · Computer Use API with browser tools for screenshot and DOM navigation; general-purpose prompts; one policy model; official grader; five independent runs.
    未知项 Exact WebArena task release and task count are not stated.
  • Public factuality datasets, no thinking · No thinking; no external tools stated; Claude Sonnet 4.5 grades Simple-QA-Verified and AA-Omniscience.
    未知项 Trial count and model sampling settings are not stated.
  • Public factuality datasets, 16K extended thinking · 16K extended-thinking budget; no external tools stated; Claude Sonnet 4.5 grades Simple-QA-Verified and AA-Omniscience. 思考预算 16K
    未知项 Trial count and model sampling settings are not stated.
  • τ²-bench Airline corrected variant · Claude Opus 4.1 simulates the user; Anthropic added a prompt addendum; corrected task setup and grading.
    未知项 Effort, trial count and detailed sampling settings are not stated.
  • Terminal-Bench 2.0 · Opus 4.6 comparison table · Terminus-2 harness; 1× guaranteed / 3× ceiling resource allocation; 5–15 samples per task across staggered batches, as stated by the comparison article.
    未知项 The article does not identify whether the Opus 4.5 value is an Anthropic reproduction or a result published by another lab, and does not state effort or the exact sample count for this row.
  • BrowseComp · comparison article · no thinking · Single agent; web search/fetch; programmatic tool calling; context compaction at 50K; 10M total-token limit; max reasoning effort; thinking disabled. max
    未知项 Trial count and exact sampling settings for the Opus 4.5 row are not stated.
  • Humanity's Last Exam · comparison article · with tools · Web search/fetch, code execution, programmatic tool calling; context compaction at 50K up to 3M total tokens; max reasoning effort; adaptive thinking enabled. max
    未知项 Trial count and exact sampling settings for the Opus 4.5 row are not stated.
  • GDPval-AA · comparison article · Artificial Analysis independent evaluation; 1,416 Elo listed for Opus 4.5 in the comparison table.
    未知项 Reasoning effort is not stated; the linked methodological details were not independently reviewed in this task.
  • MMMU-Pro · without tools · comparison article · MMMU-Pro without tools; the comparison table reports 70.6%.
    未知项 Reasoning effort, prompt format, trial count and other evaluation details are not stated in the table.
  • MMMU-Pro · with tools · comparison article · MMMU-Pro with tools; the comparison table reports 73.9%.
    未知项 Tool identity, reasoning effort, prompt format, trial count and other evaluation details are not stated in the table.
  • OpenRCA · comparison article · OpenRCA result in the comparison article; the article describes the benchmark author's harness and official grading methodology.
    未知项 Effort and trial count for the Opus 4.5 row are not stated; the evaluation was submitted for official verification, which remains pending in the article.
  • DeepSearchQA · main figure · The main figure labels the Opus 4.5 value as accuracy.
    未知项 The figure does not specify agent setup, token budget, reasoning effort or trial count; kept separate from card F1 results.

直接前代与后续发布

当前收录未确认直接前代。 官方关系依据

发布材料将 Opus 4.5 与此前型号作比较,但未在所选材料中明确唯一直接前代。

测试记录 · 31