{
  "generated": "2026-09-08",
  "benchmark": "terminal-bench",
  "results": [
    {
      "system": "Claude Opus 4.5 + Terminus 2",
      "developer": "Anthropic",
      "value": 57.8,
      "date": "2026-01-17",
      "source": {
        "url": "https://arxiv.org/abs/2601.11868",
        "kind": "paper",
        "accessed": "2026-09-04"
      },
      "conditions": {
        "split": "2.0",
        "scaffold": "Terminus 2",
        "notes": "Table 2 of the paper (Terminal-Bench 2.0)."
      }
    },
    {
      "system": "GPT-5.2 + Codex CLI",
      "developer": "OpenAI",
      "value": 62.9,
      "date": "2026-01-17",
      "source": {
        "url": "https://arxiv.org/abs/2601.11868",
        "kind": "paper",
        "accessed": "2026-09-04"
      },
      "conditions": {
        "split": "2.0",
        "scaffold": "Codex CLI",
        "notes": "Best result in Table 2 of the paper (Terminal-Bench 2.0)."
      }
    },
    {
      "system": "Opus 4.8 + Claude Code (max)",
      "developer": "Anthropic",
      "value": 23.64,
      "date": "2026-08-27",
      "source": {
        "url": "https://www.tbench.ai/leaderboard",
        "kind": "official-leaderboard",
        "accessed": "2026-09-04"
      },
      "conditions": {
        "split": "4.0",
        "scaffold": "Claude Code",
        "reasoning_effort": "max",
        "cost_usd_per_task": 19.64,
        "notes": "Leaderboard 'created_at' date; 330 trials. Cost is total run cost / 330 trials."
      }
    },
    {
      "system": "Opus 5 + Claude Code (max)",
      "developer": "Anthropic",
      "value": 51.82,
      "date": "2026-08-27",
      "source": {
        "url": "https://www.tbench.ai/leaderboard",
        "kind": "official-leaderboard",
        "accessed": "2026-09-04"
      },
      "conditions": {
        "split": "4.0",
        "scaffold": "Claude Code",
        "reasoning_effort": "max",
        "cost_usd_per_task": 18.09,
        "notes": "Leaderboard 'created_at' date; 330 trials. Cost is total run cost / 330 trials."
      }
    },
    {
      "system": "Claude Opus 5 + Claude Code",
      "developer": "Anthropic",
      "value": 30.0,
      "date": "2026-08-27",
      "source": {
        "url": "https://www.tbench.ai/news/terminal-bench-science-0-1",
        "kind": "official-leaderboard",
        "accessed": "2026-09-04"
      },
      "conditions": {
        "split": "science",
        "scaffold": "Claude Code",
        "notes": "Terminal-Bench-Science 0.1 launch post (2026-08-27); 3 trials per task over 70 tasks; total run cost $7.0k. Best model at launch. Effort level not stated."
      }
    },
    {
      "system": "GPT-5.6 Sol + Codex",
      "developer": "OpenAI",
      "value": 22.4,
      "date": "2026-08-27",
      "source": {
        "url": "https://www.tbench.ai/news/terminal-bench-science-0-1",
        "kind": "official-leaderboard",
        "accessed": "2026-09-04"
      },
      "conditions": {
        "split": "science",
        "scaffold": "Codex",
        "notes": "Terminal-Bench-Science 0.1 launch post; 3 trials per task; total run cost $4.2k. Effort level not stated."
      }
    },
    {
      "system": "Claude Fable 5 + Claude Code",
      "developer": "Anthropic",
      "value": 21.4,
      "date": "2026-08-27",
      "source": {
        "url": "https://www.tbench.ai/news/terminal-bench-science-0-1",
        "kind": "official-leaderboard",
        "accessed": "2026-09-04"
      },
      "conditions": {
        "split": "science",
        "scaffold": "Claude Code",
        "notes": "Terminal-Bench-Science 0.1 launch post; 3 trials per task; total run cost $14.2k. Effort level not stated."
      }
    },
    {
      "system": "Claude Fable 5.1 (max)",
      "developer": "Anthropic",
      "value": 52.6,
      "date": "2026-09-01",
      "source": {
        "url": "https://www.anthropic.com/claude-fable-and-mythos-5-1",
        "kind": "developer-report",
        "accessed": "2026-09-04"
      },
      "conditions": {
        "split": "science",
        "reasoning_effort": "max",
        "cost_usd_per_task": 37.9,
        "notes": "Anthropic's own harness (reproduces public Opus 5 at 29.0%, Fable 5 at 24.7%); SE +/-3.5-4.5 pts. Post dated 'September 2026', 1st used. Scaffold not stated."
      }
    },
    {
      "system": "GPT-6 Astra",
      "developer": "OpenAI",
      "value": 64.6,
      "date": "2026-09-03",
      "source": {
        "url": "https://openai.com/index/gpt-6-astra/",
        "kind": "developer-report",
        "accessed": "2026-09-04"
      },
      "conditions": {
        "split": "science",
        "notes": "OpenAI launch table (Academic section), maximum at any effort; run in OpenAI's research environment, scaffold not stated. Same table quotes Fable 5.1 at 52.6% and Opus 5 at 30.0%."
      }
    },
    {
      "system": "Fable 5.1 + Claude Code (max)",
      "developer": "Anthropic",
      "value": 57.88,
      "date": "2026-09-03",
      "source": {
        "url": "https://www.tbench.ai/leaderboard",
        "kind": "official-leaderboard",
        "accessed": "2026-09-04"
      },
      "conditions": {
        "split": "4.0",
        "scaffold": "Claude Code",
        "reasoning_effort": "max",
        "cost_usd_per_task": 18.92,
        "notes": "+/-3.8 CI over 330 trials."
      }
    },
    {
      "system": "GPT-6 Astra + Codex (max)",
      "developer": "OpenAI",
      "value": 58.18,
      "date": "2026-09-03",
      "source": {
        "url": "https://www.tbench.ai/leaderboard",
        "kind": "official-leaderboard",
        "accessed": "2026-09-04"
      },
      "conditions": {
        "split": "4.0",
        "scaffold": "Codex",
        "reasoning_effort": "max",
        "cost_usd_per_task": 9.9,
        "notes": "Rank 1 on Terminal-Bench 4.0 as of access date; +/-2.8 CI over 330 trials."
      }
    }
  ]
}