{
  "dataset": "/data/deepswe/snapshot.json",
  "datasetSha256": "af8388264d7eddf87c55b9c623c8d083e61802aca764f968fc16f07cc60130ee",
  "width": 1920,
  "figures": [
    {
      "id": "tokens",
      "variant": "full",
      "date": "2026-09-03",
      "title": "Same score, 30k to 143k tokens per task",
      "alt": "3 tied at the leading score / 70 configurations. One dot per model and reasoning effort. Tied means the pass rate rounds to the leader’s whole point.",
      "png": "tokens-2026-09.png",
      "svg": "tokens-2026-09.svg",
      "height": 2280,
      "highlights": [
        {
          "number": 1,
          "name": "GPT-6 Astra (xhigh)",
          "pass": 0.7411504424778761,
          "cost": 6.52377356460177,
          "outputTokens": 29557.327433628318,
          "steps": 28.754424778761063
        },
        {
          "number": 2,
          "name": "Gemini 3.8 Flash (high)",
          "pass": 0.738255033557047,
          "cost": 2.362349413758389,
          "outputTokens": 143242.6644295302,
          "steps": 166.3131991051454
        },
        {
          "number": 3,
          "name": "Claude Opus 5 (max)",
          "pass": 0.7364864864864865,
          "cost": 11.837583271396396,
          "outputTokens": 117565.6936936937,
          "steps": 99.0427927927928
        }
      ],
      "source": "https://deepswe.datacurve.ai/",
      "sourceSha256": "8fdbb59257d00cbb0772248bb602aafc1d20a51822388cd4eaf4625505182be7"
    },
    {
      "id": "tokens",
      "variant": "chart",
      "date": "2026-09-03",
      "title": "Same score, 30k to 143k tokens per task",
      "alt": "3 tied at the leading score / 70 configurations. One dot per model and reasoning effort. Tied means the pass rate rounds to the leader’s whole point.",
      "png": "tokens-2026-09-chart.png",
      "svg": "tokens-2026-09-chart.svg",
      "height": 1680,
      "highlights": [
        {
          "number": 1,
          "name": "GPT-6 Astra (xhigh)",
          "pass": 0.7411504424778761,
          "cost": 6.52377356460177,
          "outputTokens": 29557.327433628318,
          "steps": 28.754424778761063
        },
        {
          "number": 2,
          "name": "Gemini 3.8 Flash (high)",
          "pass": 0.738255033557047,
          "cost": 2.362349413758389,
          "outputTokens": 143242.6644295302,
          "steps": 166.3131991051454
        },
        {
          "number": 3,
          "name": "Claude Opus 5 (max)",
          "pass": 0.7364864864864865,
          "cost": 11.837583271396396,
          "outputTokens": 117565.6936936937,
          "steps": 99.0427927927928
        }
      ],
      "source": "https://deepswe.datacurve.ai/",
      "sourceSha256": "8fdbb59257d00cbb0772248bb602aafc1d20a51822388cd4eaf4625505182be7"
    },
    {
      "id": "cost",
      "variant": "full",
      "date": "2026-09-03",
      "title": "Same score, $2.36 to $11.84 per task",
      "alt": "3 tied at the leading score / 70 configurations. One dot per model and reasoning effort. Tied means the pass rate rounds to the leader’s whole point.",
      "png": "cost-2026-09.png",
      "svg": "cost-2026-09.svg",
      "height": 2280,
      "highlights": [
        {
          "number": 1,
          "name": "GPT-6 Astra (xhigh)",
          "pass": 0.7411504424778761,
          "cost": 6.52377356460177,
          "outputTokens": 29557.327433628318,
          "steps": 28.754424778761063
        },
        {
          "number": 2,
          "name": "Gemini 3.8 Flash (high)",
          "pass": 0.738255033557047,
          "cost": 2.362349413758389,
          "outputTokens": 143242.6644295302,
          "steps": 166.3131991051454
        },
        {
          "number": 3,
          "name": "Claude Opus 5 (max)",
          "pass": 0.7364864864864865,
          "cost": 11.837583271396396,
          "outputTokens": 117565.6936936937,
          "steps": 99.0427927927928
        }
      ],
      "source": "https://deepswe.datacurve.ai/",
      "sourceSha256": "8fdbb59257d00cbb0772248bb602aafc1d20a51822388cd4eaf4625505182be7"
    },
    {
      "id": "cost",
      "variant": "chart",
      "date": "2026-09-03",
      "title": "Same score, $2.36 to $11.84 per task",
      "alt": "3 tied at the leading score / 70 configurations. One dot per model and reasoning effort. Tied means the pass rate rounds to the leader’s whole point.",
      "png": "cost-2026-09-chart.png",
      "svg": "cost-2026-09-chart.svg",
      "height": 1680,
      "highlights": [
        {
          "number": 1,
          "name": "GPT-6 Astra (xhigh)",
          "pass": 0.7411504424778761,
          "cost": 6.52377356460177,
          "outputTokens": 29557.327433628318,
          "steps": 28.754424778761063
        },
        {
          "number": 2,
          "name": "Gemini 3.8 Flash (high)",
          "pass": 0.738255033557047,
          "cost": 2.362349413758389,
          "outputTokens": 143242.6644295302,
          "steps": 166.3131991051454
        },
        {
          "number": 3,
          "name": "Claude Opus 5 (max)",
          "pass": 0.7364864864864865,
          "cost": 11.837583271396396,
          "outputTokens": 117565.6936936937,
          "steps": 99.0427927927928
        }
      ],
      "source": "https://deepswe.datacurve.ai/",
      "sourceSha256": "8fdbb59257d00cbb0772248bb602aafc1d20a51822388cd4eaf4625505182be7"
    },
    {
      "id": "bill",
      "variant": "full",
      "date": "2026-09-03",
      "title": "Output tokens are under a quarter of the bill",
      "alt": "5 reasoning efforts of one model. Each bar is the mean cost of one task at the price basis DeepSWE states for Astra. The hatched part of the output segment is reasoning tokens. The dashed mark is the same tokens at OpenAI’s published rates, which carry no compute-unit line.",
      "png": "bill-2026-09.png",
      "svg": "bill-2026-09.svg",
      "height": 2448,
      "highlights": [
        {
          "effort": "low",
          "cost": 2.188762706637168,
          "atList": 1.595223439159292
        },
        {
          "effort": "medium",
          "cost": 4.379527705752213,
          "atList": 3.0754742975663714
        },
        {
          "effort": "high",
          "cost": 5.723721956194691,
          "atList": 3.9237237909292033
        },
        {
          "effort": "xhigh",
          "cost": 6.52377356460177,
          "atList": 4.429117203539823
        },
        {
          "effort": "max",
          "cost": 12.369032401327434,
          "atList": 7.497827996681416
        }
      ],
      "source": "https://deepswe.datacurve.ai/",
      "sourceSha256": "8fdbb59257d00cbb0772248bb602aafc1d20a51822388cd4eaf4625505182be7"
    },
    {
      "id": "bill",
      "variant": "chart",
      "date": "2026-09-03",
      "title": "Output tokens are under a quarter of the bill",
      "alt": "5 reasoning efforts of one model. Each bar is the mean cost of one task at the price basis DeepSWE states for Astra. The hatched part of the output segment is reasoning tokens. The dashed mark is the same tokens at OpenAI’s published rates, which carry no compute-unit line.",
      "png": "bill-2026-09-chart.png",
      "svg": "bill-2026-09-chart.svg",
      "height": 1776,
      "highlights": [
        {
          "effort": "low",
          "cost": 2.188762706637168,
          "atList": 1.595223439159292
        },
        {
          "effort": "medium",
          "cost": 4.379527705752213,
          "atList": 3.0754742975663714
        },
        {
          "effort": "high",
          "cost": 5.723721956194691,
          "atList": 3.9237237909292033
        },
        {
          "effort": "xhigh",
          "cost": 6.52377356460177,
          "atList": 4.429117203539823
        },
        {
          "effort": "max",
          "cost": 12.369032401327434,
          "atList": 7.497827996681416
        }
      ],
      "source": "https://deepswe.datacurve.ai/",
      "sourceSha256": "8fdbb59257d00cbb0772248bb602aafc1d20a51822388cd4eaf4625505182be7"
    }
  ]
}