1. 图表数据
  2. Anthropic
  3. Model Benchmark Comparison
Anthropic表格图 · 77 个数据点 · Claude Code

Model Benchmark Comparison

Anthropic 发布的一张图背后的数字，用原生图表重画。对照原图，研究数据，再做你自己的。

从图片估算 82%Claude Opus 4Claude Opus 4.6Claude Sonnet 4Anthropic 全部模型全部图表数据
图表来自 Anthropic. 来源: x.com

简化重绘

由我们用下面的数字重新绘制。悬停、点按或用方向键查看。

这是简化重绘：它用我们自己的样式展示数字，不是原图的样子。强调色标出这张图的主角。表格里带 ~ 的是估算值。

数字

77 个点中前 10 个免费
分组系列数值来源置信
Agentic terminal coding (Terminal-Bench 2.0)Opus 4.665.4%估算0.97
Agentic terminal coding (Terminal-Bench 2.0)Opus 4.559.8%估算0.97
Agentic terminal coding (Terminal-Bench 2.0)Sonnet 4.551%估算0.97
Agentic terminal coding (Terminal-Bench 2.0)Gemini 3 Pro56.2%估算0.97
Agentic terminal coding (Terminal-Bench 2.0)GPT-5 (all models)64.7%估算0.97
Agentic coding (SWE-bench Verified)Opus 4.680.8%估算0.97
Agentic coding (SWE-bench Verified)Opus 4.580.9%估算0.97
Agentic coding (SWE-bench Verified)Sonnet 4.577.2%估算0.97
Agentic coding (SWE-bench Verified)Gemini 3 Pro76.2%估算0.97
Agentic coding (SWE-bench Verified)GPT-5 (all models)80%估算0.97

Max 数据，目前留邮箱免费看

查看全部 77 个数据点并下载文件

这张图的完整表格、JSON、CSV 和 agent skill 文件。它们属于 Max；原型阶段，留邮箱即可解锁。

解锁全部 35 家巨头。每周一封，随时退订。 看看能得到什么 · 登录

Agentic computer use (OSWorld)Opus 4.672.7%估算0.97
Agentic computer use (OSWorld)Opus 4.566.3%估算0.97
Agentic computer use (OSWorld)Sonnet 4.561.4%估算0.97
Agentic tool use (τ²-bench) RetailOpus 4.691.9%估算0.97
Agentic tool use (τ²-bench) RetailOpus 4.588.9%估算0.97
Agentic tool use (τ²-bench) RetailSonnet 4.586.2%估算0.97
Agentic tool use (τ²-bench) RetailGemini 3 Pro85.3%估算0.97
Agentic tool use (τ²-bench) RetailGPT-5 (all models)82%估算0.97
Agentic tool use (τ²-bench) TelecomOpus 4.699.3%估算0.50
Agentic tool use (τ²-bench) TelecomOpus 4.598.2%估算0.50
Agentic tool use (τ²-bench) TelecomSonnet 4.598%估算0.50
Agentic tool use (τ²-bench) TelecomGemini 3 Pro98%估算0.50
Agentic tool use (τ²-bench) TelecomGPT-5 (all models)98.7%估算0.50
Scaled tool use (MCP Atlas)Opus 4.659.5%估算0.97
Scaled tool use (MCP Atlas)Opus 4.562.3%估算0.97
Scaled tool use (MCP Atlas)Sonnet 4.543.8%估算0.97
Scaled tool use (MCP Atlas)Gemini 3 Pro54.1%估算0.97
Scaled tool use (MCP Atlas)GPT-5 (all models)60.6%估算0.97
Agentic search (BrowseComp)Opus 4.684%估算0.97
Agentic search (BrowseComp)Opus 4.567.8%估算0.97
Agentic search (BrowseComp)Sonnet 4.543.9%估算0.97
Agentic search (BrowseComp)Gemini 3 Pro59.2%估算0.97
Agentic search (BrowseComp)GPT-5 (all models)77.9%估算0.97
Multidisciplinary reasoning (Humanity's Last Exam) without toolsOpus 4.640%估算0.97
Multidisciplinary reasoning (Humanity's Last Exam) without toolsOpus 4.530.8%估算0.97
Multidisciplinary reasoning (Humanity's Last Exam) without toolsSonnet 4.517.7%估算0.97
Multidisciplinary reasoning (Humanity's Last Exam) without toolsGemini 3 Pro37.5%估算0.97
Multidisciplinary reasoning (Humanity's Last Exam) without toolsGPT-5 (all models)36.6%估算0.97
Multidisciplinary reasoning (Humanity's Last Exam) with toolsOpus 4.653.1%估算0.50
Multidisciplinary reasoning (Humanity's Last Exam) with toolsOpus 4.543.4%估算0.50
Multidisciplinary reasoning (Humanity's Last Exam) with toolsSonnet 4.533.6%估算0.50
Multidisciplinary reasoning (Humanity's Last Exam) with toolsGemini 3 Pro45.8%估算0.50
Multidisciplinary reasoning (Humanity's Last Exam) with toolsGPT-5 (all models)50%估算0.50
Agentic financial analysis (Finance Agent)Opus 4.660.7%估算0.97
Agentic financial analysis (Finance Agent)Opus 4.555.9%估算0.97
Agentic financial analysis (Finance Agent)Sonnet 4.554.2%估算0.97
Agentic financial analysis (Finance Agent)Gemini 3 Pro44.1%估算0.97
Agentic financial analysis (Finance Agent)GPT-5 (all models)56.6%估算0.97
Office tasks (GDPval-AA EIo)Opus 4.61606%估算0.97
Office tasks (GDPval-AA EIo)Opus 4.51416%估算0.97
Office tasks (GDPval-AA EIo)Sonnet 4.51277%估算0.97
Office tasks (GDPval-AA EIo)Gemini 3 Pro1195%估算0.97
Office tasks (GDPval-AA EIo)GPT-5 (all models)1462%估算0.97
Novel problem-solving (ARC AGI 2)Opus 4.668.8%估算0.50
Novel problem-solving (ARC AGI 2)Opus 4.537.6%估算0.50
Novel problem-solving (ARC AGI 2)Sonnet 4.513.6%估算0.50
Novel problem-solving (ARC AGI 2)Gemini 3 Pro45.1%估算0.50
Novel problem-solving (ARC AGI 2)GPT-5 (all models)54.2%估算0.50
Graduate-level reasoning (GPQA Diamond)Opus 4.691.3%估算0.30
Graduate-level reasoning (GPQA Diamond)Opus 4.587%估算0.30
Graduate-level reasoning (GPQA Diamond)Sonnet 4.583.4%估算0.97
Graduate-level reasoning (GPQA Diamond)Gemini 3 Pro91.9%估算0.30
Graduate-level reasoning (GPQA Diamond)GPT-5 (all models)93.2%估算0.30
Visual reasoning (MMMU Pro) without toolsOpus 4.673.9%估算0.97
Visual reasoning (MMMU Pro) without toolsOpus 4.570.6%估算0.97
Visual reasoning (MMMU Pro) without toolsSonnet 4.563.4%估算0.97
Visual reasoning (MMMU Pro) without toolsGemini 3 Pro81%估算0.97
Visual reasoning (MMMU Pro) without toolsGPT-5 (all models)79.5%估算0.97
Visual reasoning (MMMU Pro) with toolsOpus 4.677.3%估算0.50
Visual reasoning (MMMU Pro) with toolsOpus 4.573.9%估算0.50
Visual reasoning (MMMU Pro) with toolsSonnet 4.568.9%估算0.50
Visual reasoning (MMMU Pro) with toolsGPT-5 (all models)80.4%估算0.50
Multilingual Q&A (MMMLU)Opus 4.691.1%估算0.97
Multilingual Q&A (MMMLU)Opus 4.590.8%估算0.97
Multilingual Q&A (MMMLU)Sonnet 4.589.5%估算0.97
Multilingual Q&A (MMMLU)Gemini 3 Pro91.8%估算0.97
Multilingual Q&A (MMMLU)GPT-5 (all models)89.6%估算0.97
下载这张图的数据

Max 订阅者可以通过 API 取同样的数据。 API 文档

JSONagent skill 文件

API · Max

用 API 调用这张图

Max 套餐可以取这张图的数字（每个数字都带来源），格式为 JSON 或 CSV，附重画规格；也可以让智能体通过 MCP 调用。图表 id：99900de8b864。

查看 Max 套餐API 文档
路径返回套餐
GET/api/v1/charts/99900de8b864标题、发布者、日期、图型、数据点数；用 Max 密钥时附数据链接开放
GET/api/v1/charts/99900de8b864/data每个数据点及其来源和置信度，加上重画规格。加 ?format=csv 得到 CSVMax
GET/api/v1/charts/99900de8b864/render?view=1本页使用的只读绘图视图开放
POST/mcp工具 get_chart_data 加这个 id，供 Claude、Cursor 等智能体使用Max
curlMax
curl -s https://viral-chart.com/api/v1/charts/99900de8b864/data \
  -H "Authorization: Bearer $VIRALCHART_KEY"
Claude CodeMax
claude mcp add --transport http viralchart https://viral-chart.com/mcp \
  --header "Authorization: Bearer $VIRALCHART_KEY"

每个密钥每分钟 60 次，超出返回 429 和 Retry-After。每个账号最多 5 个密钥，在账户页创建，只显示一次。没有 Max 时数据接口返回 402，并附套餐链接。这些数字描述的是发布者拥有的图，每个文件都带授权说明。

做一张这样的图

这张图的形式，你的数字。

实时预览。所有内容都留在你的设备上。

选择形式

简化形式：用我们自己的样式，不是原图的样子。你的数字从这张图的数据开始。

Anthropic 的更多数据

全部图表数据
从图片估算Claude 3.5 Sonnet benchmarksAnthropic · 表格图 · 34 个数据点从图片估算Software engineeringAnthropic · 柱状图 · 2 个数据点从图片估算Novel problem-solving by costAnthropic · 散点图 · 3 个数据点从图片估算相似重绘 7.4Code Arena: WebDevarena · 柱状图 · 15 个数据点从图片估算Model benchmarksAnthropic · 表格图 · 31 个数据点从图片估算相似重绘 6.0Opus 4.8Anthropic · 表格图 · 28 个数据点

常见问题

这些数字有多准？

它们由两遍独立读图估算得出，置信度 0.82（0 到 1）。图上印着的数字读得准，从坐标轴读出的值可能有几个百分点的偏差。

这些数据可以直接用吗？

这些数字描述的是 Anthropic 发布的图表。图片、版式和样式归他们所有。使用数字时请注明发布者，图片请用你自己的数据来做。

怎么拿到文件？

JSON、CSV 和 agent skill 文件在本页的邮箱解锁区。原型里它们是静态文件：/api/charts/99900de8b864.json 和 /api/charts/99900de8b864.skill.md。