{
  "benchmarks": [
    {
      "area": "agent-tasks",
      "attribution": "Berkeley Function Calling Leaderboard, UC Berkeley (Gorilla), Apache 2.0",
      "cadence": "irregular",
      "categories": [
        "inference",
        "open-weights"
      ],
      "ceiling": 100,
      "competitor": "",
      "data": "https://github.com/HuanzhiMao/BFCL-Result",
      "direction": "higher",
      "floor": 0,
      "json": "https://www.anchorterminal.com/benchmarks/bfcl/index.json",
      "licence": "Apache-2.0",
      "licenceUrl": "https://github.com/HuanzhiMao/BFCL-Result/blob/main/LICENSE",
      "measures": "Whether a model calls functions correctly: simple, parallel and multi-turn calls, knowing when no function fits, web search and memory tasks, run by the Berkeley team on every model, with native function calling and with prompting as separate rows.",
      "method": "https://gorilla.cs.berkeley.edu/leaderboard.html",
      "models": 5,
      "name": "Berkeley Function Calling Leaderboard (BFCL)",
      "newest": "2025-12-16",
      "page": "https://www.anchorterminal.com/benchmarks/bfcl/",
      "read": "2026-11-20",
      "results": 7,
      "runBy": "UC Berkeley (Gorilla)",
      "saturated": false,
      "slug": "bfcl",
      "status": "stale",
      "unit": "% overall accuracy"
    },
    {
      "area": "agent-tasks",
      "attribution": "METR, task-completion time horizons, CC BY 4.0",
      "cadence": "none (no longer updated)",
      "categories": [
        "inference",
        "open-weights"
      ],
      "competitor": "",
      "data": "https://metr.org/assets/benchmark_results_1_1.yaml",
      "direction": "higher",
      "floor": 0,
      "json": "https://www.anchorterminal.com/benchmarks/metr-time-horizon/index.json",
      "licence": "CC-BY-4.0",
      "licenceUrl": "https://creativecommons.org/licenses/by/4.0/",
      "measures": "The length of software, machine learning and security task, timed on human experts, that an agent is predicted to finish half the time, from a logistic fit over more than a hundred tasks.",
      "method": "https://metr.org/time-horizons/",
      "models": 1,
      "name": "METR time horizons",
      "newest": "2026-09-08",
      "page": "https://www.anchorterminal.com/benchmarks/metr-time-horizon/",
      "read": "2026-11-20",
      "results": 1,
      "runBy": "METR",
      "saturated": false,
      "slug": "metr-time-horizon",
      "status": "discontinued",
      "unit": "minutes at 50% success"
    },
    {
      "area": "agent-tasks",
      "attribution": "Terminal-Bench, Stanford and the Laude Institute, Apache 2.0",
      "cadence": "as models are released",
      "categories": [
        "inference",
        "open-weights",
        "agent-harnesses"
      ],
      "ceiling": 100,
      "competitor": "",
      "data": "https://github.com/harbor-framework/terminal-bench/tree/main/leaderboard/submissions",
      "direction": "higher",
      "floor": 0,
      "json": "https://www.anchorterminal.com/benchmarks/terminal-bench/index.json",
      "licence": "Apache-2.0",
      "licenceUrl": "https://github.com/harbor-framework/terminal-bench/blob/main/LICENSE",
      "measures": "An agent harness running a model in a sandboxed shell, five trials or more on every task, until the task's tests pass. A row is one harness and model pair at one reasoning effort; the model is what's ranked.",
      "method": "https://www.tbench.ai/",
      "models": 33,
      "name": "Terminal-Bench",
      "newest": "2026-11-20",
      "page": "https://www.anchorterminal.com/benchmarks/terminal-bench/",
      "read": "2026-11-20",
      "results": 51,
      "runBy": "Stanford and the Laude Institute (the Terminal-Bench maintainers)",
      "saturated": false,
      "slug": "terminal-bench",
      "status": "current",
      "unit": "% of trials passed"
    },
    {
      "area": "agent-tasks",
      "attribution": "τ²-bench, Sierra Research, MIT",
      "cadence": "as submissions land",
      "categories": [
        "inference",
        "open-weights"
      ],
      "ceiling": 100,
      "competitor": "",
      "data": "https://github.com/sierra-research/tau2-bench/tree/main/web/leaderboard/public/submissions",
      "direction": "higher",
      "floor": 0,
      "json": "https://www.anchorterminal.com/benchmarks/tau2-bench/index.json",
      "licence": "MIT",
      "licenceUrl": "https://github.com/sierra-research/tau2-bench/blob/main/LICENSE",
      "measures": "A model working as a customer-service agent under a written policy, with tools and a simulated customer, in one domain at a time (airline, retail, telecom, banking), in text or by voice. Each domain is scored on its own.",
      "method": "https://taubench.com/",
      "models": 6,
      "name": "τ²-bench",
      "newest": "2026-10-29",
      "page": "https://www.anchorterminal.com/benchmarks/tau2-bench/",
      "read": "2026-11-20",
      "results": 14,
      "runBy": "Sierra",
      "saturated": false,
      "slug": "tau2-bench",
      "status": "current",
      "unit": "% of tasks passed on the first try (pass^1)"
    },
    {
      "area": "coding",
      "attribution": "Aider polyglot leaderboard, Paul Gauthier and contributors, Apache 2.0",
      "cadence": "irregular",
      "categories": [
        "inference",
        "open-weights"
      ],
      "ceiling": 100,
      "competitor": "",
      "data": "https://github.com/Aider-AI/aider/blob/main/aider/website/_data/polyglot_leaderboard.yml",
      "direction": "higher",
      "floor": 0,
      "json": "https://www.anchorterminal.com/benchmarks/aider-polyglot/index.json",
      "licence": "Apache-2.0",
      "licenceUrl": "https://github.com/Aider-AI/aider/blob/main/LICENSE.txt",
      "measures": "225 challenging Exercism exercises in C++, Go, Java, JavaScript, Python and Rust, solved by a model inside the aider coding agent, with a second try after it sees the failing tests.",
      "method": "https://aider.chat/docs/leaderboards/",
      "models": 1,
      "name": "Aider Polyglot",
      "newest": "2025-10-03",
      "page": "https://www.anchorterminal.com/benchmarks/aider-polyglot/",
      "read": "2026-11-20",
      "results": 1,
      "runBy": "Aider (Paul Gauthier)",
      "saturated": false,
      "slug": "aider-polyglot",
      "status": "stale",
      "unit": "% of exercises solved in two tries"
    },
    {
      "area": "coding",
      "attribution": "Epoch AI, Capabilities and benchmarking (CC BY 4.0)",
      "cadence": "as models are released",
      "categories": [
        "inference",
        "open-weights"
      ],
      "ceiling": 100,
      "competitor": "",
      "data": "https://epoch.ai/data/benchmark_data.zip",
      "direction": "higher",
      "floor": 0,
      "json": "https://www.anchorterminal.com/benchmarks/mirrorcode/index.json",
      "licence": "CC BY 4.0",
      "licenceUrl": "https://epoch.ai/benchmarks/use-this-data",
      "measures": "An agent rebuilds a whole program from its behaviour alone, without the source or the internet, and must match the original's output on held-out end-to-end tests. Epoch AI and METR built it; the leaderboard runs 15 medium and large programs in two languages, three attempts each, with up to 10 billion tokens and 7 days an attempt.",
      "method": "https://epoch.ai/MirrorCode",
      "models": 19,
      "name": "MirrorCode",
      "newest": "2026-11-17",
      "page": "https://www.anchorterminal.com/benchmarks/mirrorcode/",
      "read": "2026-11-20",
      "results": 22,
      "runBy": "Epoch AI",
      "saturated": true,
      "slug": "mirrorcode",
      "status": "current",
      "unit": "% of tasks solved"
    },
    {
      "area": "coding",
      "attribution": "Epoch AI, Capabilities and benchmarking (CC BY 4.0)",
      "cadence": "as models are released",
      "categories": [
        "inference",
        "open-weights"
      ],
      "ceiling": 100,
      "competitor": "",
      "data": "https://epoch.ai/data/benchmark_data.zip",
      "direction": "higher",
      "floor": 0,
      "json": "https://www.anchorterminal.com/benchmarks/swe-bench-verified/index.json",
      "licence": "CC BY 4.0",
      "licenceUrl": "https://epoch.ai/benchmarks/use-this-data",
      "measures": "A model given a Python repository and a GitHub issue must change the code until the repository's own tests pass. Epoch AI runs 484 of the 500 human-checked issues in its own simple tool-calling scaffold, with no network, after a major scaffold upgrade in February 2026.",
      "method": "https://epoch.ai/benchmarks/swe-bench-verified",
      "models": 1,
      "name": "SWE-bench Verified",
      "newest": "2026-06-25",
      "page": "https://www.anchorterminal.com/benchmarks/swe-bench-verified/",
      "read": "2026-11-20",
      "results": 1,
      "runBy": "Epoch AI",
      "saturated": false,
      "slug": "swe-bench-verified",
      "status": "current",
      "unit": "% of issues resolved"
    },
    {
      "area": "reasoning",
      "attribution": "Epoch AI, Capabilities and benchmarking (CC BY 4.0)",
      "cadence": "as models are released",
      "categories": [
        "inference",
        "open-weights"
      ],
      "ceiling": 100,
      "competitor": "",
      "data": "https://epoch.ai/data/benchmark_data.zip",
      "direction": "higher",
      "floor": 0,
      "json": "https://www.anchorterminal.com/benchmarks/frontiermath/index.json",
      "licence": "CC BY 4.0",
      "licenceUrl": "https://epoch.ai/benchmarks/use-this-data",
      "measures": "Original research-level mathematics problems written and checked by mathematicians, answered by a Python function the model submits after reasoning with a Python tool. Epoch AI made the set and keeps most of it private; version 2 of June 2026 fixed errors in 42% of problems.",
      "method": "https://epoch.ai/benchmarks/frontiermath-tiers-1-3-v2",
      "models": 36,
      "name": "FrontierMath (Tiers 1-3)",
      "newest": "2026-11-19",
      "page": "https://www.anchorterminal.com/benchmarks/frontiermath/",
      "read": "2026-11-20",
      "results": 44,
      "runBy": "Epoch AI",
      "saturated": true,
      "slug": "frontiermath",
      "status": "current",
      "unit": "% of problems solved"
    },
    {
      "area": "reasoning",
      "attribution": "Epoch AI, Capabilities and benchmarking (CC BY 4.0)",
      "cadence": "as models are released",
      "categories": [
        "inference",
        "open-weights"
      ],
      "ceiling": 100,
      "competitor": "",
      "data": "https://epoch.ai/data/benchmark_data.zip",
      "direction": "higher",
      "floor": 0,
      "json": "https://www.anchorterminal.com/benchmarks/frontiermath-tier-4/index.json",
      "licence": "CC BY 4.0",
      "licenceUrl": "https://epoch.ai/benchmarks/use-this-data",
      "measures": "The hardest FrontierMath problems, an expansion set of 43 in version 2, each taking a specialist researcher days. Same method as FrontierMath's Tiers 1-3, run by Epoch AI on the private set.",
      "method": "https://epoch.ai/benchmarks/frontiermath-tier-4-v2",
      "models": 32,
      "name": "FrontierMath Tier 4",
      "newest": "2026-11-19",
      "page": "https://www.anchorterminal.com/benchmarks/frontiermath-tier-4/",
      "read": "2026-11-20",
      "results": 40,
      "runBy": "Epoch AI",
      "saturated": true,
      "slug": "frontiermath-tier-4",
      "status": "current",
      "unit": "% of problems solved"
    },
    {
      "area": "reasoning",
      "attribution": "Epoch AI, Capabilities and benchmarking (CC BY 4.0)",
      "cadence": "as models are released",
      "categories": [
        "inference",
        "open-weights"
      ],
      "ceiling": 100,
      "competitor": "",
      "data": "https://epoch.ai/data/benchmark_data.zip",
      "direction": "higher",
      "floor": 0,
      "json": "https://www.anchorterminal.com/benchmarks/gpqa-diamond/index.json",
      "licence": "CC BY 4.0",
      "licenceUrl": "https://epoch.ai/benchmarks/use-this-data",
      "measures": "Multiple-choice questions in biology, chemistry and physics written by PhD-level experts and hard for non-experts to answer even with the web; the Diamond subset is 198 questions. Epoch AI asks zero-shot with chain of thought and scores an answer only in its required format.",
      "method": "https://epoch.ai/benchmarks/gpqa-diamond",
      "models": 45,
      "name": "GPQA Diamond",
      "newest": "2026-11-19",
      "page": "https://www.anchorterminal.com/benchmarks/gpqa-diamond/",
      "read": "2026-11-20",
      "results": 63,
      "runBy": "Epoch AI",
      "saturated": true,
      "slug": "gpqa-diamond",
      "status": "current",
      "unit": "% of questions correct"
    },
    {
      "area": "reasoning",
      "attribution": "Epoch AI, Capabilities and benchmarking (CC BY 4.0)",
      "cadence": "as models are released",
      "categories": [
        "inference",
        "open-weights"
      ],
      "ceiling": 100,
      "competitor": "",
      "data": "https://epoch.ai/data/benchmark_data.zip",
      "direction": "higher",
      "floor": 0,
      "json": "https://www.anchorterminal.com/benchmarks/otis-mock-aime/index.json",
      "licence": "CC BY 4.0",
      "licenceUrl": "https://epoch.ai/benchmarks/use-this-data",
      "measures": "45 competition maths problems from three OTIS mock AIME exams, each with an integer answer from 0 to 999; harder than MATH Level 5 and easier than FrontierMath.",
      "method": "https://epoch.ai/benchmarks/otis-mock-aime-2024-2025",
      "models": 49,
      "name": "OTIS Mock AIME 2024-2025",
      "newest": "2026-11-19",
      "page": "https://www.anchorterminal.com/benchmarks/otis-mock-aime/",
      "read": "2026-11-20",
      "results": 68,
      "runBy": "Epoch AI",
      "saturated": true,
      "slug": "otis-mock-aime",
      "status": "current",
      "unit": "% of problems solved"
    },
    {
      "area": "reasoning",
      "attribution": "Epoch AI, Capabilities and benchmarking (CC BY 4.0)",
      "cadence": "as models are released",
      "categories": [
        "inference",
        "open-weights"
      ],
      "ceiling": 100,
      "competitor": "",
      "data": "https://epoch.ai/data/benchmark_data.zip",
      "direction": "higher",
      "floor": 0,
      "json": "https://www.anchorterminal.com/benchmarks/simpleqa-verified/index.json",
      "licence": "CC BY 4.0",
      "licenceUrl": "https://epoch.ai/benchmarks/use-this-data",
      "measures": "1,000 short factual questions with no answer choices, made by Google from OpenAI's SimpleQA. Epoch AI asks the model to guess rather than abstain, has a grader model mark each answer and reports the share answered correctly, not Google's F1 score.",
      "method": "https://epoch.ai/benchmarks/simple-qa-verified",
      "models": 30,
      "name": "SimpleQA Verified",
      "newest": "2026-11-19",
      "page": "https://www.anchorterminal.com/benchmarks/simpleqa-verified/",
      "read": "2026-11-20",
      "results": 31,
      "runBy": "Epoch AI",
      "saturated": false,
      "slug": "simpleqa-verified",
      "status": "current",
      "unit": "% of questions correct"
    },
    {
      "area": "embeddings",
      "attribution": "MTEB, embeddings-benchmark/results (CC0)",
      "cadence": "continuous (pull requests to the results repository)",
      "categories": [
        "embeddings",
        "open-weights"
      ],
      "ceiling": 100,
      "competitor": "",
      "data": "https://github.com/embeddings-benchmark/results",
      "direction": "higher",
      "floor": 62.3,
      "json": "https://www.anchorterminal.com/benchmarks/mteb-eng-v2/index.json",
      "licence": "CC0-1.0",
      "licenceUrl": "https://github.com/embeddings-benchmark/results/blob/main/LICENSE",
      "measures": "An embedding model's mean main score across the 41 English tasks of MTEB(eng, v2), covering retrieval, classification, clustering, reranking, pair classification, semantic similarity and summarisation, as the MTEB leaderboard's Mean (Task) column.",
      "method": "https://huggingface.co/spaces/mteb/leaderboard",
      "models": 11,
      "name": "MTEB English v2",
      "newest": "2026-11-20",
      "page": "https://www.anchorterminal.com/benchmarks/mteb-eng-v2/",
      "read": "2026-11-20",
      "results": 16,
      "runBy": "MTEB maintainers (an open community project started at Hugging Face), from results submitted by pull request",
      "saturated": false,
      "slug": "mteb-eng-v2",
      "status": "current",
      "unit": "mean task score (0 to 100)"
    },
    {
      "area": "embeddings",
      "attribution": "MTEB, embeddings-benchmark/results (CC0)",
      "cadence": "continuous (pull requests to the results repository)",
      "categories": [
        "embeddings",
        "open-weights"
      ],
      "ceiling": 100,
      "competitor": "",
      "data": "https://github.com/embeddings-benchmark/results",
      "direction": "higher",
      "floor": 50.81,
      "json": "https://www.anchorterminal.com/benchmarks/mteb-eng-v2-retrieval/index.json",
      "licence": "CC0-1.0",
      "licenceUrl": "https://github.com/embeddings-benchmark/results/blob/main/LICENSE",
      "measures": "An embedding model's mean nDCG@10 across the 10 retrieval tasks of MTEB(eng, v2), as the MTEB leaderboard's Retrieval column on the English tab.",
      "method": "https://huggingface.co/spaces/mteb/leaderboard",
      "models": 12,
      "name": "MTEB English v2 retrieval",
      "newest": "2026-11-20",
      "page": "https://www.anchorterminal.com/benchmarks/mteb-eng-v2-retrieval/",
      "read": "2026-11-20",
      "results": 20,
      "runBy": "MTEB maintainers (an open community project started at Hugging Face), from results submitted by pull request",
      "saturated": false,
      "slug": "mteb-eng-v2-retrieval",
      "status": "current",
      "unit": "mean nDCG@10 (0 to 100)"
    },
    {
      "area": "arena",
      "attribution": "LMArena leaderboard dataset (lmarena-ai/leaderboard-dataset), CC BY 4.0",
      "cadence": "daily",
      "categories": [
        "inference",
        "open-weights"
      ],
      "competitor": "",
      "data": "https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset",
      "direction": "higher",
      "json": "https://www.anchorterminal.com/benchmarks/lmarena-agent/index.json",
      "licence": "CC-BY-4.0",
      "licenceUrl": "https://creativecommons.org/licenses/by/4.0/",
      "measures": "How models do as agents in sessions people run through the arena, scored with inverse propensity scoring (IPS) from signals in those sessions (task outcome, bash reliability, steerability, tool hallucination, praise and complaints) rather than from pairwise votes.",
      "method": "https://arena.ai/blog/agent-arena-methodology/",
      "models": 40,
      "name": "LMArena Agent Arena",
      "newest": "2026-11-18",
      "page": "https://www.anchorterminal.com/benchmarks/lmarena-agent/",
      "read": "2026-11-20",
      "results": 46,
      "runBy": "LMArena",
      "saturated": false,
      "slug": "lmarena-agent",
      "status": "current",
      "unit": "IPS score"
    },
    {
      "area": "arena",
      "attribution": "LMArena leaderboard dataset (lmarena-ai/leaderboard-dataset), CC BY 4.0",
      "cadence": "daily",
      "categories": [
        "inference",
        "open-weights"
      ],
      "competitor": "",
      "data": "https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset",
      "direction": "higher",
      "json": "https://www.anchorterminal.com/benchmarks/lmarena-webdev/index.json",
      "licence": "CC-BY-4.0",
      "licenceUrl": "https://creativecommons.org/licenses/by/4.0/",
      "measures": "Which of two anonymous models' work people prefer on web development prompts in the arena's Code Arena, fitted as a Bradley-Terry rating.",
      "method": "https://arena.ai/blog/leaderboard-changelog/",
      "models": 53,
      "name": "LMArena Code Arena (WebDev)",
      "newest": "2026-11-19",
      "page": "https://www.anchorterminal.com/benchmarks/lmarena-webdev/",
      "read": "2026-11-20",
      "results": 58,
      "runBy": "LMArena",
      "saturated": false,
      "slug": "lmarena-webdev",
      "status": "current",
      "unit": "arena rating (Bradley-Terry)"
    },
    {
      "area": "arena",
      "attribution": "LMArena leaderboard dataset (lmarena-ai/leaderboard-dataset), CC BY 4.0",
      "cadence": "daily",
      "categories": [
        "inference",
        "open-weights"
      ],
      "competitor": "",
      "data": "https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset",
      "direction": "higher",
      "json": "https://www.anchorterminal.com/benchmarks/lmarena-text/index.json",
      "licence": "CC-BY-4.0",
      "licenceUrl": "https://creativecommons.org/licenses/by/4.0/",
      "measures": "Which of two anonymous models' answers people prefer, over the prompts they type into the arena, fitted as a Bradley-Terry rating. Rows give the raw rating and the style-controlled one, which discounts length and formatting.",
      "method": "https://arena.ai/blog/leaderboard-changelog/",
      "models": 64,
      "name": "LMArena Text Arena",
      "newest": "2026-11-19",
      "page": "https://www.anchorterminal.com/benchmarks/lmarena-text/",
      "read": "2026-11-20",
      "results": 134,
      "runBy": "LMArena",
      "saturated": false,
      "slug": "lmarena-text",
      "status": "current",
      "unit": "arena rating (Bradley-Terry)"
    },
    {
      "area": "arena",
      "attribution": "LMArena leaderboard dataset (lmarena-ai/leaderboard-dataset), CC BY 4.0",
      "cadence": "daily",
      "categories": [
        "inference",
        "open-weights"
      ],
      "competitor": "",
      "data": "https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset",
      "direction": "higher",
      "json": "https://www.anchorterminal.com/benchmarks/lmarena-vision/index.json",
      "licence": "CC-BY-4.0",
      "licenceUrl": "https://creativecommons.org/licenses/by/4.0/",
      "measures": "Which of two anonymous models' answers people prefer when the prompt comes with an image, fitted as a Bradley-Terry rating. Rows give the raw rating and the style-controlled one.",
      "method": "https://arena.ai/blog/leaderboard-changelog/",
      "models": 29,
      "name": "LMArena Vision Arena",
      "newest": "2026-11-19",
      "page": "https://www.anchorterminal.com/benchmarks/lmarena-vision/",
      "read": "2026-11-20",
      "results": 60,
      "runBy": "LMArena",
      "saturated": false,
      "slug": "lmarena-vision",
      "status": "current",
      "unit": "arena rating (Bradley-Terry)"
    }
  ],
  "independent": 637,
  "listings": 83,
  "models": 94,
  "note": "A rank is among the models we list with a result, at each model's best setting, within one version of the test. These results don't move grades.",
  "page": "https://www.anchorterminal.com/benchmarks/",
  "results": 677,
  "submitted": 40,
  "tiers": {
    "independent": "Independent",
    "ours": "Measured by Anchor Terminal",
    "self": "The maker's claim",
    "submitted": "Submitted by the maker"
  }
}
