Claude Sonnet 5 · 官方评测披露

Claude Sonnet 5 当前收录 32 项测试的官方记录。已完成列明材料的核查。

模型发布 · Claude Sonnet

官方材料与核查范围

列明主文、系统卡能力与风险章节、图表和附录已完成模型来源核对;首发历史仍需独立证据。 本轮补核主文勘误直接链接的145页系统卡,与已读146页版对照差异;常规成绩不变。 核对时间 。

当前资料不足以还原首发。 已取得首发当天官方主文回放,但尚未取得并核读全部直接附件的可靠历史正文及外链图像历史字节,无法完整还原首发材料集合;成功回放与失败材料分别保留,不用当前附件替代首发。

  • Introducing Claude Sonnet 5 · 主文 · Current page collected 2026-11-10; datePublished 2026-06-30; dateModified 2026-10-14T05:02:58Z
    公布 未确认 · 采集 · 已核查 主文全文及脚注已检查;摘要表只记录 Sonnet 5 主体成绩,Sonnet 4.6 和 Opus 4.8 参考行均未作为本发布的结果。BrowseComp 图表更新的旧数值未披露,不作推算。
  • Claude Sonnet 5 System Card · 附件 · PDF with changelog dated 2026-07-10; 146 pages; collected 2026-11-10
    公布 未确认 · 采集 · 已核查 系统卡能力、风险、安全、对齐、模型福利与附录已核读,能力图表已逐页视觉核查。只录当前主体的精确印值;材料完整不表示每项均有可排名成绩。p.20–21、23 的风险图没有精确标签;成本缩放图未标每点精确成本。生命科学图以原文 0–1 指标单列。
  • Claude Sonnet 5 System Card (145-page version directly linked by the announcement methodology correction) · 附件 · 145-page PDF with cover date June 30, 2026; no July 10 changelog or public Messages API reproduction footnote; current announcement links this version in its methodology correction. Cover date does not establish the collected file publication time.
    公布 未确认 · 采集 · 已核查 145页全文已抽取,与此前已完整核读的146页版本对照全篇内容;24处实质文本/引用编号差异逐项核查。主要是少一页7月10日changelog、BrowseComp公开API复现脚注和随之变化的引用编号;普通能力数值未变。82个嵌入图像中81个与已核读版相同,余下一项为p.80图表图例条,已实际查看p.80及BrowseComp p.122–123等差异页面;“small随”文本抽取差异不作为评测结果。复用原版全章图表核读,不把该当前可读文件当作首发快照。

配置

  • Standard configuration: adaptive thinking, max effort · Adaptive thinking at max effort; default temperature and top_p; five-trial averages unless otherwise stated. max
    未知项 Context-window size is evaluation-dependent; standard configuration is 1M tokens.
  • Xhigh effort · Xhigh effort in Terminal-Bench 2.1. xhigh
    未知项 No other model sampling details are stated for that evaluation.
  • High effort, 300K token limit · High effort with a 300K token limit. high
    未知项 Sampling parameters are not stated.
  • Extended thinking · Extended thinking enabled; no tools unless specified in result conditions.
    未知项 Specific thinking budget is not stated.
  • Thinking set to auto · Thinking set to auto; no tools.
    未知项 Exact effort selected by auto is not stated.
  • Thinking set to auto, with tools · Thinking set to auto; web search, web fetch, programmatic tool calling and code execution.
    未知项 Exact effort selected by auto is not stated.
  • Adaptive thinking, max effort, tools enabled · Adaptive thinking at max effort; tools enabled as specified per benchmark. max
  • Adaptive thinking, max effort, no tools · Adaptive thinking at max effort; no tools. max
  • Evaluation effort not specified · Only the setup explicitly named in the result's conditions is known.
    未知项 Effort and sampling may be unknown; see each result.conditions.
  • Low effort (figure label) · The figure labels this run as low effort; it does not provide a complete generation configuration. low
    未知项 Sampling parameters, token limits, and tool setup are not fully stated for the cited chart point.
  • Medium effort (figure label) · The figure labels this run as med effort; it does not provide a complete generation configuration. med
    未知项 Sampling parameters, token limits, and tool setup are not fully stated for the cited chart point.
  • High effort (figure label) · The figure labels this run as high effort; it does not provide a complete generation configuration. high
    未知项 Sampling parameters, token limits, and tool setup are not fully stated for the cited chart point.
  • Xhigh effort (figure label) · The figure labels this run as xhigh effort; it does not provide a complete generation configuration. xhigh
    未知项 Sampling parameters, token limits, and tool setup are not fully stated for the cited chart point.
  • Max effort (figure label) · The figure labels this run as max effort; it does not provide a complete generation configuration. max
    未知项 Sampling parameters, token limits, and tool setup are not fully stated for the cited chart point.
  • BrowseComp max effort, 1M-token limit · Adaptive thinking at max effort with web search, web fetch, programmatic tool calling, and code execution; 1M-token limit. max
    未知项 Whether context compaction was used at this token limit is not stated.
  • BrowseComp max effort, 3M-token limit · Adaptive thinking at max effort with web search, web fetch, programmatic tool calling, and code execution; 3M-token limit; context compaction is reported as triggering at 200K tokens when extending beyond the 1M context window. max
    未知项 Sampling parameters are not stated.

直接前代与后续发布

直接前代 Claude Sonnet 4.6。 官方关系依据

Sonnet 5 is an upgrade to Sonnet 4.6, with gains in various aspects of agentic performance.
10 项披露延续 · 22 项新增 · 21 项本次未见 · 0 项待核查
Claude Sonnet 4.6 与 Claude Sonnet 5 的披露变化
测试已核查状态Claude Sonnet 4.6Claude Sonnet 5
SWE-bench Pro · Pro新增—本次记录
Terminal-Bench · 2.1新增—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
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
GDPval-AA · v2延续

已记录的评测条件有变化

前代记录本次记录
SWE-bench Verified · Verified subset (500 problems)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
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
FrontierCode · v1新增—本次记录
CursorBench · Not stated in selected source新增—本次记录
USA Mathematical Olympiad · 2026新增—本次记录
ArXivMath · April-May 2026 releases (81 problems)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
BrowseComp · 1,266 题;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
GDP.pdf · 100 professional PDFs and prompts新增—本次记录
BenchCAD Vision2Code · 1,000-file random subset of 17,900 files新增—本次记录
ChartMuseum · 1,162 test-split questions新增—本次记录
CharXiv Reasoning · Validation split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

来源位置 main → attachment

已记录的评测条件有变化

前代记录本次记录
Legal Agent Benchmark · Public full set新增—本次记录
Toolathlon · June 2026 release used in selected source新增—本次记录
AA-Briefcase · Not stated in selected source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
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
INCLUDE · 44 languages新增—本次记录
SpatialBench Verified · Verified新增—本次记录
OfficeQA · Pro 133-question subset新增—本次记录
Real-World Finance · v2, 294 tasks新增—本次记录
SingleCellBench · LatchBio Bioinformatics single-cell RNA sequencing analysis evaluation新增—本次记录
ProteinGym Hard · System-card Hard subset; exact task count not stated新增—本次记录
Terminal-Bench · 2.0;代理框架按协议区分本次未见前代记录—
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 · Vals AI SEC-filing research benchmark本次未见前代记录—
ARC-AGI-2 · ARC Prize Foundation private dataset本次未见前代记录—
GPQA Diamond · 198-question Diamond subset本次未见前代记录—
MMMU-Pro · Standard and Vision formats本次未见前代记录—
MMMLU · 57 subjects and 14 non-English languages本次未见前代记录—
OpenRCA · 335 enterprise software failure cases本次未见前代记录—
ARC-AGI-1 · ARC Prize Foundation private dataset本次未见前代记录—
AIME · 2025 edition本次未见前代记录—
Vending-Bench · 2本次未见前代记录—
OpenAI MRCR · v2, 256K context, 8 needles本次未见前代记录—
OpenAI MRCR · v2, 1M context, 8 needles本次未见前代记录—
WebArena · Official WebArena grader, modified fuzzy_match subgrader本次未见前代记录—
WebArena-Verified · Full set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—
MedCalc-Bench · Verified dataset本次未见前代记录—
Vending-Bench Arena · Not stated in selected source本次未见前代记录—
LAB-Bench FigQA · FutureHouse FigQA; prompting and tools are recorded per result本次未见前代记录—

测试记录 · 32