[
 {
  "kind": "Kernel challenge",
  "benchmark": "Anthropic kernel challenge",
  "system": "Claude Opus 4, many hours",
  "result": "2,164 cycles",
  "value": 2164,
  "minutes": null,
  "frontier": false,
  "logo": "anthropic",
  "status": "",
  "source": "Anthropic take-home README",
  "id": "0",
  "cost": "\u2014"
 },
 {
  "kind": "Kernel challenge",
  "benchmark": "Anthropic kernel challenge",
  "system": "Claude Opus 4.5, casual Claude Code session",
  "result": "1,790 cycles",
  "value": 1790,
  "minutes": null,
  "frontier": false,
  "logo": "anthropic",
  "status": "",
  "source": "Anthropic take-home README",
  "id": "1",
  "cost": "\u2014"
 },
 {
  "kind": "Kernel challenge",
  "benchmark": "Anthropic kernel challenge",
  "system": "Claude Opus 4.5, 2 hours in its harness",
  "result": "1,579 cycles",
  "value": 1579,
  "minutes": 120,
  "frontier": true,
  "logo": "anthropic",
  "status": "",
  "source": "Anthropic take-home README",
  "id": "2",
  "cost": "\u2014"
 },
 {
  "kind": "Kernel challenge",
  "benchmark": "Anthropic kernel challenge",
  "system": "Claude Sonnet 4.5, many more than 2 hours",
  "result": "1,548 cycles",
  "value": 1548,
  "minutes": null,
  "frontier": false,
  "logo": "anthropic",
  "status": "",
  "source": "Anthropic take-home README",
  "id": "3",
  "cost": "\u2014"
 },
 {
  "kind": "Kernel challenge",
  "benchmark": "Anthropic kernel challenge",
  "system": "Claude Opus 4.5, 11.5 hours in its harness",
  "result": "1,487 cycles",
  "value": 1487,
  "minutes": 690,
  "frontier": false,
  "logo": "anthropic",
  "status": "",
  "source": "Anthropic take-home README",
  "id": "4",
  "cost": "\u2014"
 },
 {
  "kind": "Kernel challenge",
  "benchmark": "Anthropic kernel challenge",
  "system": "Claude Opus 4.5, improved harness",
  "result": "1,363 cycles",
  "value": 1363,
  "minutes": null,
  "frontier": false,
  "logo": "anthropic",
  "status": "",
  "source": "Anthropic take-home README",
  "id": "5",
  "cost": "\u2014"
 },
 {
  "kind": "Legal decisions (tax law)",
  "benchmark": "DeonticBench",
  "system": "o3",
  "result": "44.4%",
  "value": 44.4,
  "minutes": null,
  "frontier": true,
  "logo": "openai",
  "status": "",
  "source": "DeonticBench paper, Table 2 (2026-04)",
  "id": "6",
  "cost": "\u2014"
 },
 {
  "kind": "Decisions under rules",
  "benchmark": "RuleArena",
  "system": "o1-preview",
  "result": "0.47",
  "value": 47.0,
  "minutes": null,
  "frontier": true,
  "logo": "openai",
  "status": "",
  "source": "RuleArena paper, Table 3 (2024 models)",
  "id": "7",
  "cost": "\u2014"
 },
 {
  "kind": "Data analysis",
  "benchmark": "KramaBench",
  "system": "smolagents + Claude 3.7",
  "result": "55.8%",
  "value": 55.8,
  "minutes": null,
  "frontier": true,
  "logo": "anthropic",
  "status": "",
  "source": "KramaBench paper v3, Table 5 (2026-03)",
  "id": "8",
  "cost": "\u2014"
 },
 {
  "kind": "Data analysis",
  "benchmark": "KramaBench",
  "system": "Human experts",
  "result": "76.8%",
  "value": 76.8,
  "minutes": null,
  "frontier": false,
  "logo": "",
  "status": "",
  "source": "KramaBench paper v3, Table 5 (2026-03)",
  "id": "9",
  "cost": "\u2014"
 },
 {
  "kind": "Document retrieval",
  "benchmark": "BRIGHT",
  "system": "Mira-Reasoning-Retrieval",
  "result": "66.9",
  "value": 66.9,
  "minutes": null,
  "frontier": false,
  "logo": "",
  "status": "",
  "source": "BRIGHT leaderboard (2026-04-22)",
  "id": "10",
  "cost": "\u2014"
 },
 {
  "kind": "Document retrieval",
  "benchmark": "BRIGHT",
  "system": "Claude Opus 4.5 + trained retriever",
  "result": "50.9",
  "value": 50.9,
  "minutes": null,
  "frontier": true,
  "logo": "anthropic",
  "status": "",
  "source": "NVIDIA NeMo Retriever (2026-03-13)",
  "id": "11",
  "cost": "\u2014"
 },
 {
  "kind": "Finding vulnerable files",
  "benchmark": "VLoc",
  "system": "GPT-5.5 xhigh",
  "result": "0.229",
  "value": 22.9,
  "minutes": null,
  "frontier": true,
  "logo": "openai",
  "status": "",
  "source": "VLoc model-performance.json (2026-09)",
  "id": "12",
  "cost": "\u2014"
 },
 {
  "kind": "Numeric optimisation",
  "benchmark": "ALE-Bench",
  "system": "GPT-6 Astra max",
  "result": "3,164",
  "value": null,
  "minutes": null,
  "frontier": false,
  "logo": "openai",
  "status": "one plain-agent run finished, with wrong answers; OpenEvolve never ran",
  "source": "ALE-Bench leaderboard (2026-10-18)",
  "id": "13",
  "cost": "\u2014"
 },
 {
  "kind": "Testing",
  "benchmark": "SWE-Mutation",
  "system": "Claude Code + Sonnet 4.5",
  "result": "40.4%",
  "value": null,
  "minutes": null,
  "frontier": false,
  "logo": "anthropic",
  "status": "built, not run: the box was torn down",
  "source": "SWE-Mutation README (2026-07)",
  "id": "14",
  "cost": "\u2014"
 },
 {
  "kind": "Code repair",
  "benchmark": "SWE-bench Pro",
  "system": "Muse Spark 1.1",
  "result": "61.5%",
  "value": null,
  "minutes": null,
  "frontier": false,
  "logo": "",
  "status": "built, not run: the box was torn down",
  "source": "labs.scale.com",
  "id": "15",
  "cost": "\u2014"
 }
]