{
 "retrieved": "2026-11-17",
 "note": "JDI's own dated record of publicly reported AI capability results. Each row names where we saw it. Scales and versions are kept apart and are not compared with each other.",
 "gdpval": {
  "label": "Professional benchmark",
  "benchmark": "GDPval, by OpenAI",
  "what": "Defined tasks from 44 occupations in the nine sectors that contribute most to US GDP, written by professionals with an average of 14 years of experience. Experts compare the AI's deliverable with a professional's, blind.",
  "metric": "Share of comparisons where the AI deliverable was rated as good as or better than the professional's (wins or ties)",
  "points": [
   {
    "date": "2025-09-25",
    "model": "Claude Opus 4.1",
    "lab": "Anthropic",
    "v": 47.6,
    "src": "OpenAI, GDPval paper (CC BY 4.0), Sep 25, 2025",
    "url": "https://arxiv.org/abs/2510.04374",
    "grader": "expert graders"
   },
   {
    "date": "2025-12-11",
    "model": "GPT-5.2 Pro",
    "lab": "OpenAI",
    "v": 74.1,
    "src": "OpenAI, Introducing GPT-5.2, Dec 11, 2025",
    "url": "https://openai.com/index/introducing-gpt-5-2/",
    "grader": "expert human judges named for GPT-5.2 Thinking (70.9%) on the same page"
   },
   {
    "date": "2026-03-05",
    "model": "GPT-5.4",
    "lab": "OpenAI",
    "v": 83.0,
    "src": "OpenAI, Introducing GPT-5.4, Mar 5, 2026",
    "url": "https://openai.com/index/introducing-gpt-5-4/",
    "grader": "not stated"
   },
   {
    "date": "2026-04-23",
    "model": "GPT-5.5",
    "lab": "OpenAI",
    "v": 84.9,
    "src": "OpenAI, Introducing GPT-5.5, Apr 23, 2026",
    "url": "https://openai.com/index/introducing-gpt-5-5/",
    "grader": "not stated"
   }
  ],
  "scale_break": "After April 2026 OpenAI stopped publishing this wins-or-ties figure in launch posts. Later GDPval results are Elo ratings from Artificial Analysis (GDPval-AA), a different scale judged by AI models, so they are recorded separately and not joined to this line."
 },
 "gdpval_aa": {
  "label": "Professional benchmark, AI-judged rating",
  "benchmark": "GDPval-AA v2.1, by Artificial Analysis",
  "metric": "Elo rating from pairwise comparisons judged by a panel of AI models. Not a percentage.",
  "rows": [
   {
    "date": "2026-10-06",
    "model": "Claude Opus 5.5 (Max)",
    "lab": "Anthropic",
    "v": 1866,
    "src": "Klu, GDPval-AA page quoting the Artificial Analysis board as of Oct 6, 2026",
    "url": "https://klu.ai/benchmarks/gdpval-aa"
   },
   {
    "date": "2026-10-06",
    "model": "Claude Sonnet 5.5 (Max)",
    "lab": "Anthropic",
    "v": 1839,
    "src": "Klu, as above",
    "url": "https://klu.ai/benchmarks/gdpval-aa"
   },
   {
    "date": "2026-11-09",
    "model": "Claude Fable 5.2 (Max)",
    "lab": "Anthropic",
    "v": 1831,
    "src": "BenchLeader, public mirror of the board, Nov 9, 2026",
    "url": "https://www.benchleader.com/benchmarks/aa_gdpval_elo"
   }
  ]
 },
 "gdpx": {
  "label": "Working on its own",
  "benchmark": "GDPx v1.0, by SemiAnalysis",
  "what": "Multi-hour agentic tasks in 11 high-value fields, from chip design and hardware to finance, legal and software engineering. The AI works alone inside real software, with a default limit of eight hours per task. Headline score: pass@1 averaged over three attempts, each field weighted equally.",
  "metric": "Headline GDPx score, percent",
  "rows": [
   {
    "date": "2026-11-09",
    "model": "Claude Fable 5.2",
    "lab": "Anthropic",
    "v": 16.4,
    "src": "ArtificialWatch, Nov 10, 2026, reading the GDPx leaderboard",
    "url": "https://artificialwatch.com/wire/claude-fable-5-2-tops-semianalysis-gdpx-at-16-4-percent"
   },
   {
    "date": "2026-11-02",
    "model": "Claude Opus 5.5",
    "lab": "Anthropic",
    "v": 12.0,
    "src": "Unite.AI, Nov 2, 2026, on the SemiAnalysis launch",
    "url": "https://www.unite.ai/semianalysis-launches-gdpx-benchmark-spanning-11-agentic-domains/"
   },
   {
    "date": "2026-11-02",
    "model": "Gemini 4 Carbon",
    "lab": "Google",
    "v": 10.9,
    "src": "Unite.AI, Nov 2, 2026",
    "url": "https://www.unite.ai/semianalysis-launches-gdpx-benchmark-spanning-11-agentic-domains/"
   },
   {
    "date": "2026-11-02",
    "model": "GPT-6 Astra",
    "lab": "OpenAI",
    "v": 9.6,
    "src": "Unite.AI, Nov 2, 2026",
    "url": "https://www.unite.ai/semianalysis-launches-gdpx-benchmark-spanning-11-agentic-domains/"
   },
   {
    "date": "2026-11-02",
    "model": "Muse Wonder Full",
    "lab": "Meta",
    "v": 9.0,
    "src": "Unite.AI, Nov 2, 2026",
    "url": "https://www.unite.ai/semianalysis-launches-gdpx-benchmark-spanning-11-agentic-domains/"
   },
   {
    "date": "2026-11-02",
    "model": "Claude Fable 5.1",
    "lab": "Anthropic",
    "v": 8.5,
    "src": "DataStudios, Nov 2026, full launch table",
    "url": "https://www.datastudios.org/post/claude-opus-5-5-tops-semianalysis-gdpx-agent-benchmark-12-percent"
   },
   {
    "date": "2026-11-09",
    "model": "GPT-6.1 Terra",
    "lab": "OpenAI",
    "v": 8.5,
    "src": "BenchLeader, public mirror, Nov 9, 2026",
    "url": "https://www.benchleader.com/benchmarks/gdpx"
   },
   {
    "date": "2026-11-09",
    "model": "GPT-6.1 Sol",
    "lab": "OpenAI",
    "v": 8.4,
    "src": "BenchLeader, public mirror, Nov 9, 2026",
    "url": "https://www.benchleader.com/benchmarks/gdpx"
   }
  ],
  "method_src": {
   "src": "SemiAnalysis, GDPx: measuring frontier model capabilities, Nov 2, 2026",
   "url": "https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities"
  }
 },
 "horizon": {
  "label": "Autonomous task length",
  "benchmark": "Task-completion time horizon, by METR (Time Horizon 1.1)",
  "what": "The length of software and research tasks, measured by how long they take skilled humans, that an AI agent completes on its own half the time. Each point is the best model at its release date.",
  "metric": "50% time horizon, minutes of skilled-human work",
  "points": [
   {
    "date": "2023-03-14",
    "model": "GPT-4",
    "minutes": 4.0,
    "ci_low": 1.9,
    "ci_high": 8.0
   },
   {
    "date": "2023-11-06",
    "model": "GPT-4 Turbo",
    "minutes": 4.0,
    "ci_low": 1.9,
    "ci_high": 8.4
   },
   {
    "date": "2024-05-13",
    "model": "GPT-4o",
    "minutes": 7.0,
    "ci_low": 4.0,
    "ci_high": 12.9
   },
   {
    "date": "2024-06-20",
    "model": "Claude 3.5 Sonnet",
    "minutes": 11.4,
    "ci_low": 5.5,
    "ci_high": 22.4
   },
   {
    "date": "2024-09-12",
    "model": "o1-preview",
    "minutes": 20.3,
    "ci_low": 11.7,
    "ci_high": 33.4
   },
   {
    "date": "2024-10-22",
    "model": "Claude 3.5 Sonnet (new)",
    "minutes": 20.5,
    "ci_low": 10.1,
    "ci_high": 40.8
   },
   {
    "date": "2024-12-05",
    "model": "o1",
    "minutes": 38.8,
    "ci_low": 21.2,
    "ci_high": 65.0
   },
   {
    "date": "2025-02-24",
    "model": "Claude 3.7 Sonnet",
    "minutes": 60.4,
    "ci_low": 33.0,
    "ci_high": 104.2
   },
   {
    "date": "2025-04-16",
    "model": "o3",
    "minutes": 119.7,
    "ci_low": 74.6,
    "ci_high": 190.9
   },
   {
    "date": "2025-08-07",
    "model": "GPT-5",
    "minutes": 203.0,
    "ci_low": 112.6,
    "ci_high": 405.6
   },
   {
    "date": "2025-11-18",
    "model": "Gemini 3 Pro",
    "minutes": 224.3,
    "ci_low": 139.6,
    "ci_high": 379.2
   },
   {
    "date": "2025-11-24",
    "model": "Claude Opus 4.5",
    "minutes": 293.0,
    "ci_low": 161.7,
    "ci_high": 623.7
   },
   {
    "date": "2025-12-11",
    "model": "GPT-5.2",
    "minutes": 352.2,
    "ci_low": 198.1,
    "ci_high": 815.2
   },
   {
    "date": "2026-02-05",
    "model": "Claude Opus 4.6",
    "minutes": 718.8,
    "ci_low": 316.7,
    "ci_high": 3633.8
   },
   {
    "date": "2026-04-07",
    "model": "Claude Mythos Preview (early)",
    "minutes": 1044.8,
    "ci_low": 508.9,
    "ci_high": 3304.3,
    "note": "METR: measurements above 16 hours are unreliable with its current task suite."
   }
  ],
  "doubling_days": {
   "from_2023": 128.7,
   "ci_low": 104.4,
   "ci_high": 158.0,
   "all_time": 187.8
  },
  "src": "METR, Time Horizon 1.1 benchmark results file",
  "url": "https://metr.org/time-horizons/",
  "data_url": "https://metr.org/assets/benchmark_results_1_1.yaml",
  "caveat": "METR states the page is no longer actively updated (Sep 8, 2026). Its doubling estimate excludes results above 16 hours."
 }
}