Leaked AI Benchmark Report Photo

{
  "type": "photograph of a computer monitor displaying an academic technical report",
  "style": "slightly angled screen photo, visible moire pattern, LCD pixel grid, slight glare, LaTeX document formatting, serif fonts",
  "document_header": {
    "left": "4 Benchmark Evaluation",
    "right": "{argument name=\"report title\" default=\"DeepSeek-V4 Technical Report\"}"
  },
  "introductory_text": "Paragraph summarizing comprehensive evaluation of {argument name=\"main model name\" default=\"DeepSeek-V4\"} against {argument name=\"competitor model 1\" default=\"GPT-5.3\"}, {argument name=\"competitor model 2\" default=\"Claude Opus 4.6\"}, and {argument name=\"competitor model 3\" default=\"Gemini 3.1 Pro Preview\"}.",
  "visualizations": {
    "legend": "5 items with color codes: dark blue, grey, light grey, blue striped, light blue",
    "bar_charts": {
      "count": 6,
      "labels": [
        "MMLU-Pro (EM)",
        "GPQA-Diamond (Pass@1)",
        "AIME 2025 (Pass@1)",
        "LiveCodeBench (Pass@1-COT)",
        "SWE-bench Verified (Resolved)",
        "Tau-bench (Average)"
      ]
    },
    "caption": "Figure 1 | Performance comparison on core benchmarks. DeepSeek-V4 achieves state-of-the-art results across the majority of benchmarks."
  },
  "data_table": {
    "columns": [
      "Benchmark",
      "{argument name=\"main model name\" default=\"DeepSeek-V4\"}",
      "{argument name=\"competitor model 1\" default=\"GPT-5.3\"}",
      "{argument name=\"competitor model 2\" default=\"Claude Opus 4.6\"}",
      "{argument name=\"competitor model 3\" default=\"Gemini 3.1 Pro Preview\"}",
      "GPT-4.1"
    ],
    "categories": {
      "count": 4,
      "rows": [
        {"label": "General", "icon": "globe/network", "sub_items": 3},
        {"label": "Reasoning & Math", "icon": "calculator/clipboard", "sub_items": 3},
        {"label": "Code", "icon": "code brackets", "sub_items": 3},
        {"label": "Agent", "icon": "robot face", "sub_items": 3}
      ]
    }
  }
}