{
  "generated": "2026-09-08",
  "benchmark": "swe-bench-verified",
  "results": [
    {
      "system": "mini-SWE-agent + Claude Sonnet 4",
      "developer": "Anthropic",
      "value": 65.0,
      "date": "2025-07-01",
      "source": {
        "url": "https://github.com/SWE-agent/mini-swe-agent",
        "kind": "official-leaderboard",
        "accessed": "2026-09-04"
      },
      "conditions": {
        "split": "bash-only",
        "scaffold": "mini-SWE-agent"
      }
    },
    {
      "system": "Claude 4.5 Opus (high)",
      "developer": "Anthropic",
      "value": 76.8,
      "date": "2026-02-17",
      "source": {
        "url": "https://www.swebench.com/#verified",
        "kind": "official-leaderboard",
        "accessed": "2026-09-04"
      },
      "conditions": {
        "split": "bash-only",
        "scaffold": "mini-SWE-agent",
        "reasoning_effort": "high",
        "cost_usd_per_task": 0.75,
        "notes": "Highest bash-only (same-environment) entry on the official Verified leaderboard as of access date."
      }
    },
    {
      "system": "Claude Opus 4.7",
      "developer": "Anthropic",
      "value": 87.6,
      "date": "2026-04-15",
      "source": {
        "url": "https://benchlm.ai/benchmarks/swe-bench-verified",
        "kind": "aggregator",
        "accessed": "2026-09-04"
      },
      "conditions": {
        "notes": "Developer-reported number republished by aggregator; scaffold unspecified."
      }
    },
    {
      "system": "Claude Opus 5",
      "developer": "Anthropic",
      "value": 96.0,
      "date": "2026-09-03",
      "source": {
        "url": "https://benchlm.ai/benchmarks/swe-bench-verified",
        "kind": "aggregator",
        "accessed": "2026-09-04"
      },
      "conditions": {
        "notes": "Developer-reported number republished by aggregator; scaffold unspecified."
      }
    }
  ]
}