{
  "title": "DeepSeek V4.1 Flash",
  "asOf": "2026-09-10",
  "sources": {
    "ds": {
      "label": "DeepSeek · V4.1 Flash model card",
      "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash",
      "detail": "Instruct comparison, maximum reasoning effort (100). Not the base-model table."
    },
    "dsReport": {
      "label": "DeepSeek · technical report",
      "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/DeepSeek_V41_Tech_Report.pdf",
      "detail": "Tables 3–4 and §5.3. Evaluation setup and scaffold ablations."
    },
    "dsPrice": {
      "label": "DeepSeek · API prices",
      "url": "https://api-docs.deepseek.com/quick_start/pricing",
      "detail": "Peak: $0.30 input / $1.20 output. Off-peak: $0.15 / $0.60. V4 Pro: $1.32 / $3.96 peak; half off-peak."
    },
    "opus": {
      "label": "Anthropic · Opus 5 launch",
      "url": "https://www.anthropic.com/news/claude-opus-5",
      "detail": "Launch comparison image: DeepSWE v1.1 68.8%; HLE without tools 56.3%."
    },
    "sonnet": {
      "label": "Anthropic · Sonnet 5 launch",
      "url": "https://www.anthropic.com/news/claude-sonnet-5",
      "detail": "Launch comparison image: Terminal-Bench 2.1 80.4%; HLE without tools 43.2%."
    },
    "fable": {
      "label": "Anthropic · Fable 5.1 launch",
      "url": "https://www.anthropic.com/claude-fable-and-mythos-5-1",
      "detail": "Terminal-Bench 4.0 55.8%; HLE without tools 60.9%. Some safeguard interventions fall back to other Claude models."
    },
    "claudePrice": {
      "label": "Anthropic · API prices",
      "url": "https://platform.claude.com/docs/en/about-claude/pricing",
      "detail": "Standard base input and output, not Fast mode or batch."
    },
    "gpt": {
      "label": "OpenAI · GPT-5.6 launch",
      "url": "https://openai.com/index/gpt-5-6/",
      "detail": "Coding and academic tables. Single-agent Sol, Terra and Luna; exclude Sol Ultra."
    },
    "gpt6": {
      "label": "OpenAI · GPT-6 Astra launch",
      "url": "https://openai.com/index/gpt-6-astra/",
      "detail": "Coding and academic tables. The comparison also reports competitor measurements."
    },
    "gptPrice": {
      "label": "OpenAI · API prices",
      "url": "https://developers.openai.com/api/docs/pricing",
      "detail": "Standard short-context pricing: Sol $4 / $20; Terra $2 / $12; Luna $0.20 / $1.20; Astra $10 / $50. Batch and Flex are separate, half-price tiers."
    },
    "google": {
      "label": "Google DeepMind · Gemini launch table",
      "url": "https://deepmind.google/models/gemini/",
      "detail": "Gemini 3.8 Flash: Terminal-Bench 2.1 89.4%; DeepSWE v1.1 73.7%; Terminal-Bench 4.0 19.1%."
    },
    "googleMethod": {
      "label": "Google DeepMind · evaluation methods",
      "url": "https://deepmind.google/models/evals-methodology/gemini-3-8-flash",
      "detail": "DeepSWE uses mini-SWE with high thinking. Terminal-Bench 2.1 uses Terminus 2. Terminal-Bench 4.0 comes from its public leaderboard."
    },
    "googlePrice": {
      "label": "Google · Gemini API prices",
      "url": "https://ai.google.dev/gemini-api/docs/pricing#gemini-3.8-flash",
      "detail": "Standard introductory $0.75 / $3.75 through 31 Dec 2026; $1.50 / $7.50 from 1 Jan 2027."
    },
    "kimi": {
      "label": "Moonshot · Kimi K3 model card",
      "url": "https://huggingface.co/moonshotai/Kimi-K3",
      "detail": "Max effort. GPQA 93.5%; Terminal-Bench 2.1 88.3%; DeepSWE 67.5% in Kimi Code (67.3% in mini-SWE); HLE full without tools 43.5%."
    },
    "kimiPrice": {
      "label": "Moonshot · Kimi K3 API prices",
      "url": "https://platform.kimi.ai/docs/pricing/chat-k3",
      "detail": "Cache-miss input $3; output $15 per million tokens."
    },
    "glm": {
      "label": "Z.ai · GLM-5.3 model card",
      "url": "https://huggingface.co/zai-org/GLM-5.3",
      "detail": "Terminal-Bench 2.1 88.2% in Claude Code; DeepSWE v1.1 66.9% in mini-SWE."
    },
    "glmFlash": {
      "label": "Z.ai · GLM-5.3 Flash model card",
      "url": "https://huggingface.co/zai-org/GLM-5.3-Flash",
      "detail": "Launch chart: Terminal-Bench 2.1 84.3%; DeepSWE v1.1 63.4%. Footnotes specify Claude Code and mini-SWE respectively."
    },
    "glmPrice": {
      "label": "Z.ai · API prices",
      "url": "https://docs.z.ai/guides/overview/pricing",
      "detail": "GLM-5.3 $1.40 / $4.40; GLM-5.3 Flash $0.15 / $0.50 per million tokens."
    }
  },
  "benchmarks": {
    "tb21": {
      "name": "Terminal-Bench 2.1",
      "short": "Terminal 2.1",
      "unit": "Pass@1 (%)",
      "description": "Agents completing tasks in a terminal.",
      "note": "Same benchmark version, different agent harnesses. DeepSeek uses DSH Minimal, Kimi uses Kimi Code, GLM uses Claude Code, and Gemini uses Terminus 2. Do not read this as a controlled model-only test."
    },
    "swe": {
      "name": "DeepSWE v1.1",
      "short": "DeepSWE",
      "unit": "Resolved (%)",
      "description": "Long-horizon software engineering.",
      "note": "DeepSeek uses mini-SWE, not its own harness, for the 74.2% headline. Kimi uses Kimi Code. Anthropic’s Opus launch reports 68.8%, versus 74.0% in DeepSeek’s table. A 0.1–0.5 point lead is not evidence of a reliable win."
    },
    "tb4": {
      "name": "Terminal-Bench 4.0",
      "short": "Terminal 4.0",
      "unit": "Pass@1 (%)",
      "description": "Harder, newer terminal tasks.",
      "note": "Do not merge with Terminal-Bench 2.1 or 3.0. Fable 5.1 can fall back to other Claude models after safeguard interventions. GPT-6 Astra is outside DeepSeek’s comparison table."
    },
    "gpqa": {
      "name": "GPQA Diamond",
      "short": "GPQA",
      "unit": "Pass@1 (%)",
      "description": "Graduate-level science reasoning.",
      "note": "No agentic composite here. DeepSeek reports 90.9%. Google’s 95.3% and Fable’s 93.7% here come from OpenAI’s comparison, not their own launch pages. Differences in prompts and reasoning budgets remain."
    },
    "hle": {
      "name": "Humanity’s Last Exam",
      "short": "HLE",
      "unit": "Pass@1 (%)",
      "description": "Full-set reasoning, without tools.",
      "note": "Only full-set, no-tool results. Exclude DeepSeek V4 Pro and GLM-5.3: their launch-table scores are text-only. Also exclude HLE-Verified, tool-assisted HLE, and missing scores rather than treating them as zero."
    }
  },
  "models": [
    {
      "id": "flash",
      "name": "DeepSeek V4.1 Flash",
      "short": "V4.1 Flash",
      "lab": "DeepSeek",
      "input": 0.3,
      "output": 1.2,
      "priceSource": "dsPrice",
      "offPeak": true,
      "focus": true,
      "priceNote": "Peak by default. Off-peak halves both rates. reasoning_effort=100."
    },
    {
      "id": "glmflash",
      "name": "GLM-5.3 Flash",
      "short": "GLM Flash",
      "lab": "Z.ai",
      "input": 0.15,
      "output": 0.5,
      "priceSource": "glmPrice"
    },
    {
      "id": "luna",
      "name": "GPT-5.6 Luna",
      "short": "GPT Luna",
      "lab": "OpenAI",
      "input": 0.2,
      "output": 1.2,
      "priceSource": "gptPrice"
    },
    {
      "id": "gemini",
      "name": "Gemini 3.8 Flash",
      "short": "Gemini 3.8",
      "lab": "Google",
      "input": 0.75,
      "output": 3.75,
      "priceSource": "googlePrice",
      "priceNote": "Introductory rates through 31 Dec 2026; double from 1 Jan 2027."
    },
    {
      "id": "glm",
      "name": "GLM-5.3",
      "short": "GLM-5.3",
      "lab": "Z.ai",
      "input": 1.4,
      "output": 4.4,
      "priceSource": "glmPrice"
    },
    {
      "id": "dspro",
      "name": "DeepSeek V4 Pro",
      "short": "V4 Pro",
      "lab": "DeepSeek",
      "input": 1.32,
      "output": 3.96,
      "priceSource": "dsPrice",
      "offPeak": true,
      "priceNote": "Current API version V4-Pro-0813. Retires on 14 Sept 2026; calls then route to V4.1 Flash."
    },
    {
      "id": "terra",
      "name": "GPT-5.6 Terra",
      "short": "GPT Terra",
      "lab": "OpenAI",
      "input": 2,
      "output": 12,
      "priceSource": "gptPrice"
    },
    {
      "id": "sonnet",
      "name": "Claude Sonnet 5",
      "short": "Sonnet 5",
      "lab": "Anthropic",
      "input": 2,
      "output": 10,
      "priceSource": "claudePrice"
    },
    {
      "id": "kimi",
      "name": "Kimi K3",
      "short": "Kimi K3",
      "lab": "Moonshot",
      "input": 3,
      "output": 15,
      "priceSource": "kimiPrice"
    },
    {
      "id": "sol",
      "name": "GPT-5.6 Sol",
      "short": "GPT Sol",
      "lab": "OpenAI",
      "input": 4,
      "output": 20,
      "priceSource": "gptPrice",
      "priceNote": "Direct Standard short-context rate. $2 / $10 in the fetched direct pricing table is Batch/Flex, not Standard. Reseller offers may differ."
    },
    {
      "id": "opus",
      "name": "Claude Opus 5",
      "short": "Opus 5",
      "lab": "Anthropic",
      "input": 5,
      "output": 25,
      "priceSource": "claudePrice"
    },
    {
      "id": "astra",
      "name": "GPT-6 Astra",
      "short": "GPT-6 Astra",
      "lab": "OpenAI",
      "input": 10,
      "output": 50,
      "priceSource": "gptPrice"
    },
    {
      "id": "fable",
      "name": "Claude Fable 5.1",
      "short": "Fable 5.1",
      "lab": "Anthropic",
      "input": 10,
      "output": 50,
      "priceSource": "claudePrice",
      "priceNote": "Some published scores use fallback Claude models on safeguard interventions."
    }
  ],
  "rows": [
    {
      "model": "flash",
      "benchmark": "tb21",
      "value": 90.6,
      "source": "ds",
      "kind": "deepseek",
      "note": "Maximum effort (100)."
    },
    {
      "model": "flash",
      "benchmark": "swe",
      "value": 74.2,
      "source": "ds",
      "kind": "deepseek",
      "note": "Maximum effort (100)."
    },
    {
      "model": "flash",
      "benchmark": "tb4",
      "value": 31.2,
      "source": "ds",
      "kind": "deepseek",
      "note": "Maximum effort (100)."
    },
    {
      "model": "flash",
      "benchmark": "gpqa",
      "value": 90.9,
      "source": "ds",
      "kind": "deepseek",
      "note": "Maximum effort (100)."
    },
    {
      "model": "flash",
      "benchmark": "hle",
      "value": 36.8,
      "source": "ds",
      "kind": "deepseek",
      "note": "Maximum effort (100)."
    },
    {
      "model": "opus",
      "benchmark": "tb21",
      "value": 89.1,
      "source": "ds",
      "kind": "deepseek",
      "note": "As reported by DeepSeek; not necessarily measured in the same harness."
    },
    {
      "model": "opus",
      "benchmark": "swe",
      "value": 74,
      "source": "ds",
      "kind": "deepseek",
      "note": "As reported by DeepSeek; not necessarily measured in the same harness."
    },
    {
      "model": "opus",
      "benchmark": "tb4",
      "value": 51.8,
      "source": "ds",
      "kind": "deepseek",
      "note": "As reported by DeepSeek; not necessarily measured in the same harness."
    },
    {
      "model": "opus",
      "benchmark": "gpqa",
      "value": 93.4,
      "source": "ds",
      "kind": "deepseek",
      "note": "As reported by DeepSeek; not necessarily measured in the same harness."
    },
    {
      "model": "opus",
      "benchmark": "hle",
      "value": 56.3,
      "source": "ds",
      "kind": "deepseek",
      "note": "As reported by DeepSeek; not necessarily measured in the same harness."
    },
    {
      "model": "sol",
      "benchmark": "tb21",
      "value": 88.8,
      "source": "ds",
      "kind": "deepseek",
      "note": "As reported by DeepSeek; not necessarily measured in the same harness."
    },
    {
      "model": "sol",
      "benchmark": "swe",
      "value": 73,
      "source": "ds",
      "kind": "deepseek",
      "note": "As reported by DeepSeek; not necessarily measured in the same harness."
    },
    {
      "model": "sol",
      "benchmark": "tb4",
      "value": 39.9,
      "source": "ds",
      "kind": "deepseek",
      "note": "As reported by DeepSeek; not necessarily measured in the same harness."
    },
    {
      "model": "sol",
      "benchmark": "gpqa",
      "value": 94.1,
      "source": "ds",
      "kind": "deepseek",
      "note": "As reported by DeepSeek; not necessarily measured in the same harness."
    },
    {
      "model": "sol",
      "benchmark": "hle",
      "value": 44.5,
      "source": "ds",
      "kind": "deepseek",
      "note": "As reported by DeepSeek; not necessarily measured in the same harness."
    },
    {
      "model": "kimi",
      "benchmark": "tb21",
      "value": 88.3,
      "source": "ds",
      "kind": "deepseek",
      "note": "As reported by DeepSeek; not necessarily measured in the same harness."
    },
    {
      "model": "kimi",
      "benchmark": "swe",
      "value": 67.5,
      "source": "ds",
      "kind": "deepseek",
      "note": "As reported by DeepSeek; not necessarily measured in the same harness."
    },
    {
      "model": "kimi",
      "benchmark": "tb4",
      "value": 12.6,
      "source": "ds",
      "kind": "deepseek",
      "note": "As reported by DeepSeek; not necessarily measured in the same harness."
    },
    {
      "model": "kimi",
      "benchmark": "gpqa",
      "value": 92.9,
      "source": "ds",
      "kind": "deepseek",
      "note": "As reported by DeepSeek; not necessarily measured in the same harness."
    },
    {
      "model": "kimi",
      "benchmark": "hle",
      "value": 43.5,
      "source": "ds",
      "kind": "deepseek",
      "note": "As reported by DeepSeek; not necessarily measured in the same harness."
    },
    {
      "model": "glm",
      "benchmark": "tb21",
      "value": 88.2,
      "source": "ds",
      "kind": "deepseek",
      "note": "As reported by DeepSeek; not necessarily measured in the same harness."
    },
    {
      "model": "glm",
      "benchmark": "swe",
      "value": 66.9,
      "source": "ds",
      "kind": "deepseek",
      "note": "As reported by DeepSeek; not necessarily measured in the same harness."
    },
    {
      "model": "glm",
      "benchmark": "tb4",
      "value": 37.9,
      "source": "ds",
      "kind": "deepseek",
      "note": "As reported by DeepSeek; not necessarily measured in the same harness."
    },
    {
      "model": "glm",
      "benchmark": "gpqa",
      "value": 88.1,
      "source": "ds",
      "kind": "deepseek",
      "note": "As reported by DeepSeek; not necessarily measured in the same harness."
    },
    {
      "model": "dspro",
      "benchmark": "tb21",
      "value": 87.9,
      "source": "ds",
      "kind": "deepseek",
      "note": "As reported by DeepSeek; not necessarily measured in the same harness."
    },
    {
      "model": "dspro",
      "benchmark": "swe",
      "value": 62.7,
      "source": "ds",
      "kind": "deepseek",
      "note": "As reported by DeepSeek; not necessarily measured in the same harness."
    },
    {
      "model": "dspro",
      "benchmark": "tb4",
      "value": 12.4,
      "source": "ds",
      "kind": "deepseek",
      "note": "As reported by DeepSeek; not necessarily measured in the same harness."
    },
    {
      "model": "dspro",
      "benchmark": "gpqa",
      "value": 92.4,
      "source": "ds",
      "kind": "deepseek",
      "note": "As reported by DeepSeek; not necessarily measured in the same harness."
    },
    {
      "model": "flash",
      "benchmark": "tb21",
      "value": 90.6,
      "source": "ds",
      "kind": "own",
      "note": "DSH Minimal; 1M context; max effort; N=3 per task."
    },
    {
      "model": "flash",
      "benchmark": "swe",
      "value": 74.2,
      "source": "ds",
      "kind": "own",
      "note": "mini-SWE; 1M context; max effort; N=8 per task."
    },
    {
      "model": "flash",
      "benchmark": "tb4",
      "value": 31.2,
      "source": "ds",
      "kind": "own",
      "note": "DSH Minimal; 1M context; max effort."
    },
    {
      "model": "flash",
      "benchmark": "gpqa",
      "value": 90.9,
      "source": "ds",
      "kind": "own",
      "note": "Max effort; instruct model."
    },
    {
      "model": "flash",
      "benchmark": "hle",
      "value": 36.8,
      "source": "ds",
      "kind": "own",
      "note": "Full set, no tools. Text-only result is 39.1%, not used here."
    },
    {
      "model": "glmflash",
      "benchmark": "tb21",
      "value": 84.3,
      "source": "glmFlash",
      "kind": "own",
      "note": "Claude Code 2.1.207; 6-hour timeout; max effort."
    },
    {
      "model": "glmflash",
      "benchmark": "swe",
      "value": 63.4,
      "source": "glmFlash",
      "kind": "own",
      "note": "mini-SWE; 400K context; 6-hour timeout."
    },
    {
      "model": "luna",
      "benchmark": "tb21",
      "value": 84.7,
      "source": "gpt",
      "kind": "own",
      "note": "Single-agent launch result, not Ultra."
    },
    {
      "model": "luna",
      "benchmark": "swe",
      "value": 67.2,
      "source": "gpt",
      "kind": "own",
      "note": "Single-agent launch result, not Ultra."
    },
    {
      "model": "luna",
      "benchmark": "gpqa",
      "value": 92.3,
      "source": "gpt",
      "kind": "own",
      "note": "Single-agent launch result, not Ultra."
    },
    {
      "model": "terra",
      "benchmark": "tb21",
      "value": 87.4,
      "source": "gpt",
      "kind": "own",
      "note": "Single-agent launch result, not Ultra."
    },
    {
      "model": "terra",
      "benchmark": "swe",
      "value": 69.6,
      "source": "gpt",
      "kind": "own",
      "note": "Single-agent launch result, not Ultra."
    },
    {
      "model": "terra",
      "benchmark": "gpqa",
      "value": 92.9,
      "source": "gpt",
      "kind": "own",
      "note": "Single-agent launch result, not Ultra."
    },
    {
      "model": "sol",
      "benchmark": "tb21",
      "value": 88.8,
      "source": "gpt",
      "kind": "own",
      "note": "Single-agent launch result, not Ultra."
    },
    {
      "model": "sol",
      "benchmark": "swe",
      "value": 72.7,
      "source": "gpt",
      "kind": "own",
      "note": "Single-agent launch result, not Ultra."
    },
    {
      "model": "sol",
      "benchmark": "gpqa",
      "value": 94.6,
      "source": "gpt",
      "kind": "own",
      "note": "Single-agent launch result, not Ultra."
    },
    {
      "model": "sol",
      "benchmark": "tb4",
      "value": 37.3,
      "source": "gpt6",
      "kind": "own",
      "note": "OpenAI’s later Astra comparison; DeepSeek reports 39.9%."
    },
    {
      "model": "sol",
      "benchmark": "hle",
      "value": 44.5,
      "source": "ds",
      "kind": "other",
      "note": "Full-set no-tool result from DeepSeek; also matches Kimi’s launch table."
    },
    {
      "model": "terra",
      "benchmark": "tb4",
      "value": 23.6,
      "source": "google",
      "kind": "other",
      "note": "Google launch comparison / Terminal-Bench public leaderboard."
    },
    {
      "model": "gemini",
      "benchmark": "tb21",
      "value": 89.4,
      "source": "google",
      "kind": "own",
      "note": "Terminus 2; see Google evaluation methods."
    },
    {
      "model": "gemini",
      "benchmark": "swe",
      "value": 73.7,
      "source": "google",
      "kind": "own",
      "note": "mini-SWE; high thinking. OpenAI’s Astra table reports 73.8%."
    },
    {
      "model": "gemini",
      "benchmark": "tb4",
      "value": 19.1,
      "source": "google",
      "kind": "own",
      "note": "Google cites the official leaderboard, highest reported thinking level."
    },
    {
      "model": "gemini",
      "benchmark": "gpqa",
      "value": 95.3,
      "source": "gpt6",
      "kind": "other",
      "note": "OpenAI’s Astra comparison; not a Google self-report."
    },
    {
      "model": "glm",
      "benchmark": "tb21",
      "value": 88.2,
      "source": "glm",
      "kind": "own",
      "note": "Claude Code 2.1.207; 6-hour timeout."
    },
    {
      "model": "glm",
      "benchmark": "swe",
      "value": 66.9,
      "source": "glm",
      "kind": "own",
      "note": "mini-SWE; 400K context; 6-hour timeout."
    },
    {
      "model": "glm",
      "benchmark": "tb4",
      "value": 37.9,
      "source": "ds",
      "kind": "other",
      "note": ""
    },
    {
      "model": "glm",
      "benchmark": "gpqa",
      "value": 88.1,
      "source": "ds",
      "kind": "other",
      "note": ""
    },
    {
      "model": "dspro",
      "benchmark": "tb21",
      "value": 87.9,
      "source": "ds",
      "kind": "own",
      "note": "DeepSeek’s own V4 Pro comparison; API currently serves V4-Pro-0813."
    },
    {
      "model": "dspro",
      "benchmark": "swe",
      "value": 62.7,
      "source": "ds",
      "kind": "own",
      "note": "DeepSeek’s own V4 Pro comparison; API currently serves V4-Pro-0813."
    },
    {
      "model": "dspro",
      "benchmark": "tb4",
      "value": 12.4,
      "source": "ds",
      "kind": "own",
      "note": "DeepSeek’s own V4 Pro comparison; API currently serves V4-Pro-0813."
    },
    {
      "model": "dspro",
      "benchmark": "gpqa",
      "value": 92.4,
      "source": "ds",
      "kind": "own",
      "note": "DeepSeek’s own V4 Pro comparison; API currently serves V4-Pro-0813."
    },
    {
      "model": "sonnet",
      "benchmark": "tb21",
      "value": 80.4,
      "source": "sonnet",
      "kind": "own",
      "note": ""
    },
    {
      "model": "sonnet",
      "benchmark": "hle",
      "value": 43.2,
      "source": "sonnet",
      "kind": "own",
      "note": "Full set, no tools."
    },
    {
      "model": "sonnet",
      "benchmark": "swe",
      "value": 53.8,
      "source": "google",
      "kind": "other",
      "note": "Google launch comparison, citing Datacurve leaderboard."
    },
    {
      "model": "sonnet",
      "benchmark": "tb4",
      "value": 12.4,
      "source": "google",
      "kind": "other",
      "note": ""
    },
    {
      "model": "kimi",
      "benchmark": "tb21",
      "value": 88.3,
      "source": "kimi",
      "kind": "own",
      "note": "Kimi Code harness, max effort."
    },
    {
      "model": "kimi",
      "benchmark": "swe",
      "value": 67.5,
      "source": "kimi",
      "kind": "own",
      "note": "Kimi Code harness; mini-SWE result is 67.3%."
    },
    {
      "model": "kimi",
      "benchmark": "gpqa",
      "value": 93.5,
      "source": "kimi",
      "kind": "own",
      "note": "Max effort. DeepSeek reports 92.9%."
    },
    {
      "model": "kimi",
      "benchmark": "hle",
      "value": 43.5,
      "source": "kimi",
      "kind": "own",
      "note": "HLE-Full without tools."
    },
    {
      "model": "kimi",
      "benchmark": "tb4",
      "value": 12.6,
      "source": "ds",
      "kind": "other",
      "note": ""
    },
    {
      "model": "opus",
      "benchmark": "tb21",
      "value": 89.1,
      "source": "ds",
      "kind": "other",
      "note": "Also appears in Google’s launch table; not found on Opus’s launch page."
    },
    {
      "model": "opus",
      "benchmark": "swe",
      "value": 68.8,
      "source": "opus",
      "kind": "own",
      "note": "Anthropic launch image. DeepSeek reports 74.0%; OpenAI’s Astra comparison reports 73.7%."
    },
    {
      "model": "opus",
      "benchmark": "tb4",
      "value": 52.3,
      "source": "fable",
      "kind": "own",
      "note": "Anthropic’s Fable 5.1 comparison; DeepSeek reports 51.8%."
    },
    {
      "model": "opus",
      "benchmark": "gpqa",
      "value": 93.4,
      "source": "ds",
      "kind": "other",
      "note": "DeepSeek comparison; OpenAI’s Astra comparison reports 93.7%."
    },
    {
      "model": "opus",
      "benchmark": "hle",
      "value": 56.3,
      "source": "opus",
      "kind": "own",
      "note": "Opus launch image, full set without tools. Later Fable 5.1 comparison reports 56.6%."
    },
    {
      "model": "astra",
      "benchmark": "swe",
      "value": 74.1,
      "source": "gpt6",
      "kind": "own",
      "note": ""
    },
    {
      "model": "astra",
      "benchmark": "tb4",
      "value": 57.9,
      "source": "gpt6",
      "kind": "own",
      "note": ""
    },
    {
      "model": "astra",
      "benchmark": "gpqa",
      "value": 96,
      "source": "gpt6",
      "kind": "own",
      "note": ""
    },
    {
      "model": "fable",
      "benchmark": "swe",
      "value": 67.4,
      "source": "gpt6",
      "kind": "other",
      "note": "OpenAI’s comparison. Fable may fall back after safeguard interventions."
    },
    {
      "model": "fable",
      "benchmark": "tb4",
      "value": 55.8,
      "source": "fable",
      "kind": "own",
      "note": "Includes fallback models on certain safeguard interventions."
    },
    {
      "model": "fable",
      "benchmark": "gpqa",
      "value": 93.7,
      "source": "gpt6",
      "kind": "other",
      "note": "OpenAI’s comparison, not an Anthropic launch self-report."
    },
    {
      "model": "fable",
      "benchmark": "hle",
      "value": 60.9,
      "source": "fable",
      "kind": "own",
      "note": "Full set without tools; biological safeguards may trigger fallback."
    }
  ],
  "scaffolds": {
    "launch": {
      "name": "Launch best",
      "tb21": 90.6,
      "swe": 74.2
    },
    "claude": {
      "name": "Claude Code",
      "tb21": 88,
      "swe": 69.8
    },
    "pi": {
      "name": "Pi",
      "tb21": 86.1,
      "swe": 66.2
    },
    "codex": {
      "name": "Codex",
      "tb21": 84.1,
      "swe": 65.6
    },
    "opencode": {
      "name": "OpenCode",
      "tb21": 85,
      "swe": 65.5
    },
    "mini": {
      "name": "mini-SWE",
      "tb21": 90.3,
      "swe": 74.2
    }
  }
}