Benchmark Atlas

排名与横向对比 · 厂商发布

Master Benchmark Table

将 coding、reasoning、agentic search 等异质指标放进一张可扫描总表,保留测试条件脚注。

tablemulti-benchmarkrelease
Frontier model benchmark sheet 将 coding、reasoning、agentic search 等异质指标放进一张可扫描总表,保留测试条件脚注。 示意数据,仅复现信息结构,不代表最新榜单结果。 Frontier model benchmark sheet Anthropic / Claude DEMO DATA Opus 4.6Opus 4.5GPT-5.xGemini ProCODINGSWE-bench82.178.476.874.2AGENTTerminal-Bench65.858.961.457.1REASONGPQA Diamond91.388.190.287.4SEARCHBrowseComp84.776.986.280.1conditions · tools on where applicable · normalized display
Illustrative demo data · renderer: masterTable

Information grammar

它如何组织信息

分组行 + 主模型描边 + 单元格双层高亮

Visual system
VS-01 Warm editorial table
Component ID
claude-master-table

Use when

适用判断

适合模型发布总览;不适合指标口径差异过大且无法写清脚注的场景。

Source lineage

Anthropic / Claude

暖纸、珊瑚主系列、强表格与误差区间

Open formal source ↗

Structured data

可替换的数据模型

示意数据,仅复现信息结构,不代表最新榜单结果。

{
  "columns": [
    "Opus 4.6",
    "Opus 4.5",
    "GPT-5.x",
    "Gemini Pro"
  ],
  "rows": [
    {
      "group": "CODING",
      "label": "SWE-bench",
      "values": [
        82.1,
        78.4,
        76.8,
        74.2
      ]
    },
    {
      "group": "AGENT",
      "label": "Terminal-Bench",
      "values": [
        65.8,
        58.9,
        61.4,
        57.1
      ]
    },
    {
      "group": "REASON",
      "label": "GPQA Diamond",
      "values": [
        91.3,
        88.1,
        90.2,
        87.4
      ]
    },
    {
      "group": "SEARCH",
      "label": "BrowseComp",
      "values": [
        84.7,
        76.9,
        86.2,
        80.1
      ]
    }
  ]
}

Related components

同一图表家族的其他语法