benchmark_id,name,short_name,category,homepage_url,description,license_note,display_order,measurement_count,primary_measurement_count,model_count,visible_model_count,exact_model_count,contamination_warning_model_count,other_supplemental_model_count,latest_freshness_at
epoch-eci,Epoch Capabilities Index,ECI,capability-index,https://epoch.ai/eci,Composite capability index fitted by Epoch AI across more than 50 benchmarks.,"Official Epoch AI aggregate dataset, CC-BY-4.0.",1,87,86,86,86,86,0,0,2026-10-30
aa-intelligence-index,Artificial Analysis Intelligence Index,AA Index,capability-index,https://artificialanalysis.ai/leaderboards/models,"Artificial Analysis Intelligence Index v5.0 (October 2026): nine evaluations across four equally weighted categories (Knowledge Work 25%, Coding 25%, Science 25%, General 25%): AA-Briefcase v1.1, AutomationBench-AA, GDP.pdf, FrontierCode 2.0, Terminal-Bench 4.0, Terminal-Bench-Science 0.1, CritPt v2 (Hard Subset), AA-Omniscience v1.1 and HLE-Diamond. One value per model and reasoning setting.",Aggregate index values with attribution to Artificial Analysis (republication agreement 2026-09-02); the underlying evaluations retain their upstream terms.,2,70,28,28,28,28,0,0,2026-10-30T01:23:12Z
arena-text,Arena Text,Arena,preference,https://arena.ai/leaderboard/text,Human preference ranking for text models.,Leaderboard dataset: CC-BY-4.0,3,94,79,79,79,79,0,0,2026-10-28
livebench,LiveBench,LiveBench,general,https://livebench.ai/,Frequently refreshed objective evaluation across broad capabilities.,,4,61,57,57,57,57,0,0,2026-10-13T00:37:05Z
hle,Humanity's Last Exam · Full multimodal,HLE Full,expert-knowledge,https://labs.scale.com/leaderboard/humanitys_last_exam,"Comparable values use Scale AI Labs' uniform finalized 2,500-question full-multimodal, no-tools evaluation, including row-specific dates and confidence intervals. Artificial Analysis' text-only run is a different question set and has its own column; other text-only, tool-assisted and vendor-specific runs remain supplemental and never replace this lane.",Aggregate leaderboard facts are attributed to Scale AI Labs; gated benchmark questions are never redistributed.,5,144,16,54,54,16,0,38,2026-10-23T21:21:02Z
hle-text,Humanity's Last Exam · Text-only,HLE Text,expert-knowledge,https://artificialanalysis.ai/,"Artificial Analysis' uniform text-only run: 2,158 of the 2,500 finalized questions (May 2025 revision), no tools, pass@1 with an equality-checker judge. Kept apart from the full multimodal lane because the question set differs; the two columns must not be read against each other.",Aggregate leaderboard facts are attributed to Artificial Analysis; gated benchmark questions are never redistributed.,6,156,64,64,64,64,0,0,2026-10-29T10:53:25Z
simpleqa-verified,SimpleQA Verified,SimpleQA,factuality,https://www.kaggle.com/benchmarks/deepmind/simpleqa-verified,Short-answer factual knowledge and abstention quality.,Aggregate scores from Epoch AI (CC-BY-4.0) and the official Kaggle leaderboard with attribution to Kaggle / Google DeepMind (confirmation 2026-09-02).,7,98,62,66,66,62,0,4,2026-10-26T20:00:37.000Z
gpqa-diamond,GPQA Diamond,GPQA,science,https://www.vals.ai/benchmarks/gpqa,"Vals AI's independent comparison on all 198 GPQA Diamond questions. Every canonical row combines the same zero-shot and five-shot chain-of-thought prompt lanes; Epoch's, Kaggle's and Artificial Analysis' independent no-tools runs over the same 198 questions share the comparable lane, vendor protocols remain separate supplemental measurements.",Only minimal aggregate score facts are reproduced with Vals AI attribution. The GPQA dataset is CC-BY-4.0; the Vals leaderboard page remains copyrighted.,8,321,52,92,92,52,0,40,2026-10-26T20:00:37.000Z
mmlu-pro,MMLU-Pro,MMLU-Pro,knowledge,https://www.vals.ai/benchmarks/mmlu_pro,Broad academic multiple-choice knowledge and reasoning across fourteen subjects; Vals AI's independent run over the MMLU-Pro set.,Only minimal aggregate score facts are reproduced with Vals AI attribution; MMLU-Pro questions are not redistributed.,9,70,53,62,62,53,0,9,2026-09-02T19:01:55Z
otis-mock-aime,OTIS Mock AIME 2024–2025,OTIS AIME,mathematics,https://epoch.ai/benchmarks/otis-mock-aime-2024-2025,"Epoch AI's independent evaluation on 45 challenging olympiad-style mathematics problems, with the real run date retained per result.","Official Epoch AI aggregate results, CC-BY-4.0.",10,136,78,78,78,78,0,0,2026-10-26T20:00:37.000Z
frontiermath,FrontierMath T1–3,FrontierMath T1–3,mathematics,https://epoch.ai/benchmarks/frontiermath-tiers-1-3-v2,Independent Epoch AI runs on the 285 private problems from the 295-problem FrontierMath Tiers 1–3 v2 set; the 10 public examples and Tier 4 are not mixed into this column.,"Official Epoch AI aggregate results, CC-BY-4.0; benchmark questions are not redistributed.",11,81,67,67,67,67,0,0,2026-10-26T20:00:37.000Z
arc-agi-2,ARC-AGI-2,ARC-AGI-2,abstraction,https://epoch.ai/benchmarks/arc-agi-2,"Abstract rule induction over 360 evaluation tasks, reported as pass@2 accuracy with reasoning variants kept separate.","Aggregate leaderboard scores from ARC Prize's verified board and Epoch AI's mirror with attribution (ARC Prize confirmation 2026-09-02, CC-BY-4.0 mirror); benchmark tasks are not redistributed.",12,385,54,54,54,54,0,0,2026-10-28T21:25:02Z
frontiermath-tier-4,FrontierMath T4,FrontierMath T4,mathematics,https://epoch.ai/benchmarks/frontiermath-tier-4-v2,Independent Epoch AI runs on the 41 private research-level problems from the 43-problem FrontierMath Tier 4 v2 set; the two public examples and Tiers 1–3 are not mixed into this column.,"Official Epoch AI aggregate results, CC-BY-4.0; benchmark questions are not redistributed.",13,59,52,52,52,52,0,0,2026-09-22T22:14:05.000Z
mmmu-pro,MMMU-Pro,MMMU-Pro,multimodal,https://www.vals.ai/benchmarks/mmmu,"Multimodal expert knowledge and reasoning; Vals AI's independent MMMU-Pro run is the comparable lane, official and vendor rows remain supplemental.",Only minimal aggregate score facts are reproduced with Vals AI attribution; MMMU-Pro items are not redistributed.,14,180,42,70,70,42,0,28,2026-10-29T10:53:25Z
critpt,CritPt,CritPt,science,https://epoch.ai/benchmarks/critpt,"Research-level physics reasoning: 71 open-ended challenges decomposed into about 190 machine-verifiable checkpoints, graded automatically; Artificial Analysis runs the evaluation and Epoch AI mirrors it, both share the comparable lane.","Aggregate leaderboard scores from Epoch AI with attribution, CC-BY-4.0; benchmark problems are not redistributed.",15,333,88,88,88,88,0,0,2026-10-29T17:25:13Z
ifbench,IFBench,IFBench,instruction-following,https://github.com/allenai/IFBench,"Precise instruction following on 300 prompts with 344 verifiable, out-of-domain constraints; instruction-level strict accuracy as measured by the Swallow LLM Leaderboard (final answer only, rule-based checkers).","Aggregate leaderboard scores from the Swallow LLM Leaderboard with attribution, CC-BY-4.0.",16,8,8,8,8,8,0,0,2026-09-30T18:53:19Z
ifeval,IFEval,IFEval,instruction-following,https://github.com/google-research/google-research/tree/master/instruction_following_eval,Instruction following. Provider reports and subset studies retain their scoring and sample-set labels; they do not form a uniform cross-model ranking.,,16,7,0,4,3,0,0,3,2026-07-10
simplebench,SimpleBench,SimpleBench,reasoning,https://epoch.ai/benchmarks/simplebench,Trick-question common-sense and spatial reasoning where human baselines still beat frontier models; average accuracy over five runs as published on the SimpleBench leaderboard.,"Aggregate leaderboard scores from Epoch AI with attribution, CC-BY-4.0; benchmark questions are not redistributed.",16,42,36,36,36,36,0,0,2026-09-02T09:26:45Z
chess-puzzles,Chess Puzzles,Chess,games,https://epoch.ai/benchmarks/chess-puzzles,"Epoch AI's 100 novel, engine-generated chess puzzles with exactly one best next move; evaluated by Epoch AI with a run timestamp per score.","Aggregate scores from Epoch AI with attribution, CC-BY-4.0.",17,113,64,64,64,64,0,0,2026-10-19T22:39:00.000Z
arc-agi-1,ARC-AGI-1,ARC-AGI-1,abstraction,https://epoch.ai/benchmarks/arc-agi,Original ARC-AGI abstraction and reasoning puzzles on the semi-private evaluation set as reported by ARC Prize; every score keeps its reported reasoning configuration.,"Aggregate leaderboard scores from Epoch AI with attribution, CC-BY-4.0; ARC Prize tasks are not redistributed.",18,201,52,52,52,52,0,0,2026-10-28T21:25:03Z
swe-bench-verified,SWE-bench Verified,SWE-bench,coding-agent,https://www.vals.ai/benchmarks/swebench,"Resolution rate from Vals AI's independent runs over all 500 SWE-bench Verified tasks with one documented minimal bash-only mini-swe-agent harness. Epoch's 484-task v2 run and vendor configurations remain separate supplemental lanes and are never mixed into this ranking. Vals AI has archived this benchmark (last update 1 September 2026) and no longer evaluates new model releases, nor does the official leaderboard (last entry 26 February 2026); newer models are no longer added here.",Only minimal aggregate score facts are reproduced with Vals AI attribution. Epoch supplemental results remain CC-BY-4.0; SWE-bench task data is not redistributed.,19,107,56,60,60,56,0,4,2026-09-02T19:01:44Z
swe-rebench-v2,SWE-rebench v2,SWE-rebench,coding-agent,https://swe-rebench.com/,Fresh time-windowed software issue resolution.,The underlying SWE-rebench leaderboard dataset is CC-BY-4.0; displayed scores retain source attribution and exact window metadata.,20,47,11,35,35,5,6,24,2026-08-30T11:02:42Z
swe-bench-live,SWE-bench-Live · Lite,SWE-Live,coding-agent,https://swe-bench-live.github.io/,"Continuously refreshed software issue resolution with the agent, model and submitted reasoning label retained per run.",Aggregate leaderboard results with attribution to SWE-bench-Live (maintainers' confirmation 2026-09-02); benchmark and harness MIT.,21,9,6,6,6,6,0,0,2026-10-08
deepswe,DeepSWE,DeepSWE,coding-agent,https://epoch.ai/benchmarks/deepswe,"Original long-horizon software engineering tasks across 91 active repositories; model, harness and reasoning effort remain attached to every score.",Aggregate scores from Epoch AI with attribution; DeepSWE repository Apache-2.0.,22,154,28,35,35,28,0,7,2026-10-13
swe-bench-pro-v2,SWE-bench Pro V2 · Scale,SWE-Pro V2,coding-agent,https://labs.scale.com/leaderboard/swe_bench_pro_public_v2,"Scale AI Labs' SWE-Bench Pro V2: the refreshed public split of 642 professional software-engineering tasks across 11 repositories (89 invalid tasks dropped, contradictory instructions corrected) under a locked evaluation protocol in which the agent reaches only the model endpoint and every diff is re-graded on a clean image. Each score keeps its agent harness, reasoning effort, confidence interval and row date. Successor of the retired 731-task public leaderboard; its values are not comparable with V2.",Aggregate leaderboard facts with Scale AI Labs attribution; benchmark repository terms apply.,23,16,16,16,16,16,0,0,2026-10-16T19:17:34Z
cursorbench,CursorBench,CursorBench,coding-agent,https://epoch.ai/benchmarks/cursorbench,"Coding-agent evaluation with explicit reasoning levels and per-task cost, token and step metadata; Epoch AI's mirror of the CursorBench leaderboard is the comparable lane, attributed provider reports remain supplemental.",Aggregate leaderboard scores from Epoch AI's CC-BY-4.0 mirror with attribution to Epoch AI and Cursor (rights holders' confirmation 2026-09-02); benchmark tasks are not redistributed.,24,123,31,31,31,31,0,0,2026-10-28T21:25:01Z
gso,GSO,GSO,coding-agent,https://epoch.ai/benchmarks/gso,Software performance-optimisation tasks in real repositories; OPT@1 success rate with the reported scaffold and evaluation date.,"Aggregate leaderboard scores from Epoch AI with attribution, CC-BY-4.0; benchmark tasks are not redistributed.",25,23,17,17,17,17,0,0,2026-09-27
frontiercode,FrontierCode 1.1 · Main,FrontierCode,coding-agent,https://epoch.ai/benchmarks/frontiercode,"Hard coding-agent tasks graded for mergeability over the 100-task Main subset, retaining harness and reasoning effort.",Aggregate leaderboard scores from Epoch AI's CC-BY-4.0 mirror with attribution (Epoch AI confirmation 2026-09-02); benchmark tasks are not redistributed.,26,40,40,40,40,40,0,0,2026-10-28T21:25:01Z
ioi,IOI,IOI,coding,https://www.vals.ai/benchmarks/ioi,International Olympiad in Informatics 2024 and 2025 problems evaluated by Vals AI with olympiad graders; overall score.,Only minimal aggregate score facts are reproduced with Vals AI attribution; problems are not redistributed.,27,41,39,39,39,39,0,0,2026-10-21T19:24:33Z
livecodebench,LiveCodeBench,LiveCodeBench,coding,https://www.vals.ai/benchmarks/lcb,"Competitive programming problems released after model training cutoffs; Vals AI's independent implementation, overall pass@1. Vals AI has archived this benchmark (last update 1 September 2026) and no longer evaluates new model releases; newer models are no longer added here.",Only minimal aggregate score facts are reproduced with Vals AI attribution; problems are not redistributed.,28,64,52,60,60,52,0,8,2026-09-02T19:01:50Z
scicode,SciCode,SciCode,coding,https://epoch.ai/benchmarks/scicode,"Scientific research coding measured by subproblem accuracy with executable Python test cases; Artificial Analysis runs the evaluation and Epoch AI mirrors it, both share the comparable lane.","Aggregate leaderboard scores from Epoch AI with attribution, CC-BY-4.0; benchmark tasks are not redistributed.",29,307,77,84,84,77,0,7,2026-10-29T10:53:25Z
ale-bench,ALE-Bench,ALE-Bench,coding,https://epoch.ai/benchmarks/ale-bench,Score-based algorithmic optimisation contests from AtCoder Heuristic Contests; performance rating across problems as published on the ALE-Bench leaderboard.,"Aggregate leaderboard scores from Epoch AI with attribution, CC-BY-4.0; contest problems are not redistributed.",31,75,65,65,65,65,0,0,2026-10-28T21:25:04Z
vibe-code-bench,Vibe Code Bench,Vibe Code,coding-agent,https://www.vals.ai/benchmarks/vibe-code,"Building complete web applications from scratch in an agent harness, evaluated by Vals AI; overall score.",Only minimal aggregate score facts are reproduced with Vals AI attribution.,32,81,77,77,77,77,0,0,2026-10-28T17:21:40Z
webdev-arena,Text Arena · Coding,Text Arena,coding,https://epoch.ai/benchmarks/webdev-arena,"Human preference rating for generated web applications in React, TypeScript and Tailwind.","Aggregate leaderboard scores from Epoch AI and LMArena with attribution, CC-BY-4.0.",33,193,86,86,86,86,0,0,2026-10-28
mirrorcode,MirrorCode,MirrorCode,coding-agent,https://epoch.ai/benchmarks/mirrorcode,"Long-horizon agent benchmark for reimplementing complete programs: 15 targets, two languages and three attempts per task.","Official Epoch AI aggregate results, CC-BY-4.0; task content is not redistributed.",34,15,14,14,14,14,0,0,2026-10-21T16:49:08.000Z
programbench,ProgramBench,ProgramBench,coding-agent,https://www.vals.ai/benchmarks/programbench,"Vals AI's independent run over the 200 public ProgramBench tasks: rebuild a command-line program from scratch, given only its execute-only binary and a behavioural specification, with one uniform harness (mini-swe-agent). Share of tasks whose rebuild passes all hidden tests (Fully Resolved).",Only minimal aggregate score facts are reproduced with Vals AI attribution; benchmark tasks are not redistributed.,34,55,53,53,53,53,0,0,2026-10-21T23:21:15Z
weirdml-v3,WeirdML v3,WeirdML v3,coding-agent,https://epoch.ai/benchmarks/weirdml-v3,"Successor of WeirdML v2 with a new task set: 11 hand-made machine-learning and data-analysis tasks with unfamiliar data, solved agentically inside coding agents such as Codex CLI, Claude Code, Gemini CLI, Kimi Code or OpenCode. The official score rewards token efficiency as well as final performance. Not comparable with WeirdML v2.","Aggregate leaderboard scores from Epoch AI with attribution, CC-BY-4.0; benchmark tasks are not redistributed.",35,25,25,25,25,25,0,0,2026-10-29T17:25:05Z
terminal-bench-2.1,Terminal-Bench 2.1,Terminal 2.1,terminal-agent,https://www.vals.ai/benchmarks/terminal-bench-2-1,Vals AI's independent run over Terminal-Bench 2.1 with one uniform harness; attributed current vendor reports remain supplemental.,Only minimal aggregate score facts are reproduced with Vals AI attribution; benchmark tasks are not redistributed.,37,171,57,69,69,57,0,12,2026-10-26T22:05:28Z
terminal-bench-4,Terminal-Bench 4.0,Terminal 4,terminal-agent,https://www.vals.ai/benchmarks/terminal-bench-4,"Vals AI's independent run over all 66 Terminal-Bench 4.0 tasks with one uniform harness (mini-swe-agent, avg@3); Artificial Analysis' step-limited run and attributed vendor reports remain supplemental.",Only minimal aggregate score facts are reproduced with Vals AI attribution; benchmark tasks are not redistributed.,38,143,47,59,59,47,0,12,2026-10-29T17:25:13Z
sre-bench,SRE Bench,SRE Bench,terminal-agent,https://www.vals.ai/benchmarks/srebench,"Software reverse engineering (not site reliability): Vals AI's run over 262 contamination-controlled binaries built from 19 in-house programs, partly with anti-analysis protection; every model works in the same shell-only agent loop with the same toolkit. Share of instances with all six tasks solved (Fully Solved).",Only minimal aggregate score facts are reproduced with Vals AI attribution; the proprietary benchmark tasks are not redistributed.,39,31,31,31,31,31,0,0,2026-10-21T23:21:15Z
tau3-banking-aa,τ³-Banking · Artificial Analysis harness,τ³-Bank AA,multi-turn-agent,https://artificialanalysis.ai/,"Artificial Analysis' uniform re-run of the banking domain: 97 tasks, tau2-bench tools with BM25/grep retrieval, a GPT-5.4 Mini user simulator and assertion judge, pass@1 over five repeats. The user simulator and retrieval setup change the measurement, so this column must not be read against the official harness.",Aggregate leaderboard facts are attributed to Artificial Analysis; benchmark task content is never redistributed.,40,88,43,43,43,43,0,0,2026-10-08T22:23:24Z
mcp-atlas,MCP-Atlas · Scale,MCP-Atlas,tool-use,https://labs.scale.com/leaderboard/mcp_atlas,"Real-world multi-step tool use across 1,000 tasks, 36 MCP servers and 220 tools. The comparable lane uses Scale's April 2026 rescored protocol, all public and private tasks, a 100-tool-call budget and the exact row-specific publication date.",Aggregate leaderboard facts with Scale AI Labs attribution; benchmark dataset and harness retain their upstream terms.,41,24,23,24,24,23,0,1,2026-09-17T14:59:54Z
skillsbench,SkillsBench v1.1,SkillsBench,tool-use,https://skillsbench.ai/,Audited agent benchmark across 87 professional tasks. The comparable value uses supplied skills; the paired no-skills run remains available as a separate protocol.,"Official release-aligned leaderboard dataset, Apache-2.0.",42,43,15,22,22,15,0,7,2026-08-12
apex-agents,APEX-Agents,APEX,tool-use,https://epoch.ai/benchmarks/apex-agents,"Long-horizon professional work across 480 banking, consulting and legal tasks using documents, office applications and code execution.",Aggregate scores from Epoch AI with attribution; benchmark tasks and worlds CC-BY.,43,75,63,63,63,63,0,0,2026-10-28T21:25:02Z
bfcl-v4,Berkeley Function Calling Leaderboard V4,BFCL V4,tool-use,https://gorilla.cs.berkeley.edu/leaderboard,"Function calling, multi-turn tool use and hallucination resistance.","Official BFCL V4 aggregate leaderboard, benchmark and evaluation code Apache-2.0.",44,16,6,10,10,6,0,4,2026-08-28T22:03:18Z
arena-agent,LMArena Agent Arena,Agent Arena,multi-turn-agent,https://arena.ai/leaderboard/agent,"Human preference evaluation of agent outcomes. Every score retains the submitted model, disclosed thinking label, confidence interval and observation count.","Official revision-pinned LMArena leaderboard aggregate, CC-BY-4.0.",45,44,43,43,43,43,0,0,2026-10-28
appworld,AppWorld,AppWorld,tool-use,https://github.com/StonyBrookNLP/appworld-leaderboard,"Interactive coding-agent benchmark across controllable applications and APIs; model, method, scaffold and task/scenario metrics remain attached to each run.",Aggregate server-evaluated scores from the official revision-pinned leaderboard with attribution to Stony Brook NLP (confirmation 2026-09-02).,46,5,5,5,5,5,0,0,2026-09-02
vending-bench-2,Vending-Bench 2,Vending 2,multi-turn-agent,https://epoch.ai/benchmarks/vending-bench-2,Long-horizon business simulation by Andon Labs: mean final balance in USD after running a vending machine business over a simulated year.,"Aggregate leaderboard scores from Epoch AI with attribution, CC-BY-4.0.",47,56,51,51,51,51,0,0,2026-10-29T17:25:04Z
gdpval-aa-v2.1,GDPval-AA v2.1,GDPval-AA v2.1,professional-work,https://artificialanalysis.ai/evaluations/gdpval-aa,"Artificial Analysis' independent agentic evaluation on 220 OpenAI GDPval tasks across 44 occupations, scored as an Elo rating from blind pairwise comparisons of complete work products. Version 2.1 refits the Elo scale (Crowd-BT, DeepSeek V4.1 Flash (max) pinned at 1600 instead of human experts at 1000), so its values are not comparable with GDPval-AA v2; the Artificial Analysis leaderboard is the comparable lane.",Aggregate score facts with attribution to Artificial Analysis (republication agreement 2026-09-02); OpenAI's GDPval dataset retains its upstream terms.,48,170,79,79,79,79,0,0,2026-10-30T01:23:12Z
terminal-bench-3,Terminal-Bench 3.0,Terminal 3,terminal-agent,https://hub.harborframework.com/datasets/terminal-bench/terminal-bench/latest?tab=leaderboard&leaderboard=3-0-0,"Official agent-plus-model leaderboard over 74 hard terminal tasks; agent, effort and trial count stay attached to every score.",Official public Harbor leaderboard; benchmark dataset Apache-2.0.,51,11,11,11,11,11,0,0,2026-08-19T03:41:59.486409+00:00
terminal-bench-science-0.1,Terminal-Bench-Science 0.1,TB-Science,terminal-agent,https://www.vals.ai/benchmarks/terminal-bench-science,"Vals AI's independent run over the 70 Terminal-Bench-Science 0.1 research-workflow tasks with one uniform harness (mini-swe-agent, pass@1); rows on another harness and attributed vendor reports remain supplemental.",Only minimal aggregate score facts are reproduced with Vals AI attribution; benchmark tasks are not redistributed.,52,43,36,36,36,36,0,0,2026-10-26T22:05:28Z
gdpval,GDPval,GDPval,professional-work,https://epoch.ai/benchmarks/gdpval,OpenAI GDPval: win rate of model deliverables against industry experts across 44 occupations as published on the official leaderboard.,"Aggregate leaderboard scores from Epoch AI with attribution, CC-BY-4.0; GDPval tasks are not redistributed.",53,4,4,4,4,4,0,0,2026-09-02T09:27:05Z
osworld-2,OSWorld 2.0,OSWorld 2,computer-use-agent,https://github.com/xlang-ai/OSWorld-V2,Computer-use agent benchmark over 108 long-horizon real-world desktop workflows. Partial-credit and strict-completion scores stay as separate protocol splits tied to an exact benchmark release.,"Only attributed aggregate score facts are stored; gated task assets, environments and other benchmark artifacts remain under their upstream terms.",54,18,9,10,10,9,0,1,2026-09-02T21:27:40Z
automationbench,AutomationBench,Automation,tool-use,https://github.com/zapier/AutomationBench,Programmatically graded cross-application business workflows spanning 47 simulated SaaS tools. Public-task and private official-leaderboard protocols must remain separate.,"Only attributed aggregate score facts are stored; the public task repository, private leaderboard and associated artifacts retain their upstream terms.",55,3,0,3,3,0,0,3,2026-10-13
tau3-banking,τ³-Banking · Official harness,τ³-Banking,multi-turn-agent,https://taubench.com/leaderboard/,"Knowledge-intensive multi-turn banking support tasks with retrieval, tools and a simulated user, as published on the official leaderboard (standard tau3 tool set, all-tools retrieval). Evaluations run with a different user simulator or retrieval setup are a different measurement and are kept in their own column.",,57,26,22,25,25,22,0,3,2026-08-03
weirdml,WeirdML,WeirdML,coding,https://epoch.ai/benchmarks/weirdml,Unusual machine-learning tasks that models solve end-to-end by writing and running code; accuracy across the current WeirdML task set.,"Aggregate leaderboard scores from Epoch AI with attribution, CC-BY-4.0; benchmark tasks are not redistributed.",58,77,49,49,49,49,0,0,2026-10-08T22:23:14Z