Datasets:

Meridal
/

releves

id,name,creator,year,measures,items,format,unit,human_baseline,saturation,known_issues,url
aa_intelligence_index,Artificial Analysis Intelligence Index v4.3.2,Artificial Analysis,2026,"One synthetic score that averages ten independent tests of agents, coding, scientific reasoning and general knowledge.","10 evaluations (AA-Briefcase, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, AA-LCR, AA-Omniscience, HLE, GDP.pdf, CritPt)","Composite index (weighted average: Agents 30%, Coding 20%, Scientific reasoning 20%, General 30%)",index points,,"Not saturated: top model 57.6 (Claude Opus 5.5, max effort) on 2026-10-06, unchanged on 2026-10-08.","Version changes often (v4.0 Jan 2026 to v4.3.2 Sep-Oct 2026); scores are not comparable across versions. Each reasoning-effort setting is scored separately. English, mostly text.",https://artificialanalysis.ai/methodology/intelligence-benchmarking
epoch_capabilities_index,Epoch Capabilities Index (ECI),Epoch AI,,"One capability score fitted across dozens of benchmarks, so that models tested on different benchmarks can be compared on one scale.",,Statistical composite (item-response-style fit over many benchmarks),ECI points (with 90% CI),,Not saturated (open scale).,"Depends on which benchmarks each model has results for; very new models (e.g. Gemini 4 Argon, MiMo-V2.6-Pro, Mistral Large 4) not yet scored.",https://epoch.ai/benchmarks
lmarena_text,LMArena Text Arena (overall),LMArena,,"Which answer people prefer in blind side-by-side chats, turned into a rating.","8,626,731 votes across 413 models (snapshot of 2026-10-02)",Pairwise human preference votes,Elo-style score,,"Not saturated, but the top 20 models sit within about 40 points with overlapping confidence intervals.","Measures preference, not correctness; sensitive to style and length (style control applied by default); new models have few votes and wide intervals (marked preliminary). The site now redirects from lmarena.ai to arena.ai ('formerly known as LMArena'); no new snapshot after 2026-10-02 as of 2026-10-08.",https://lmarena.ai/leaderboard/text
gpqa_diamond,GPQA Diamond,"David Rein, Samuel R. Bowman et al. (GPQA paper)",2023,"Very hard graduate-level multiple-choice questions in biology, physics and chemistry that cannot be answered by simple web search.",198,Multiple choice (4 options),% accuracy,"PhD-level experts 65% (74% discounting clear mistakes); skilled non-experts with web access 34% (on the full GPQA set, per the paper abstract).","Near saturation: best scores around 95-96% (Epoch AI, Artificial Analysis). Removed from the AA Intelligence Index in v4.2.","Small set (198 questions), so differences of 1-2 points are noise; the paper notes that experts reach 74% once clear mistakes they identified are discounted, so a few answers are disputable and the practical ceiling is below 100%.",https://arxiv.org/abs/2311.12022
hle,Humanity's Last Exam (HLE),Center for AI Safety and Scale AI,2025,"Expert-written closed-ended questions across dozens of academic subjects, designed to be the hardest academic exam for AI.","2,500 public questions (Artificial Analysis evaluates 2,158)","Exact-match short answers and multiple choice (24% MC, 14% need an image)",% accuracy,,"Not saturated: about 61% without tools (Claude Opus 5.5, AA); vendors report up to 67.7% with tools.","With-tools and no-tools scores are mixed in announcements; answer grading uses an LLM judge (judge choice differs between AA, Scale and vendors); CAIS released an updated ""HLE-Rolling"" set on 2026-09-17.",https://lastexam.ai/
frontiermath_t1_3,"FrontierMath Tiers 1-3 (v2, private)",Epoch AI,2026,Unpublished research-level mathematics problems with automatically checkable answers.,,Numeric / symbolic answers checked automatically,% accuracy,,Approaching saturation for the very top (93.7% for GPT-6 Astra and GPT-6.1 Sol).,"Held privately and run only by Epoch AI, so results cannot be reproduced by third parties.",https://epoch.ai/benchmarks
frontiermath_t4,"FrontierMath Tier 4 (v2, private)",Epoch AI,2026,The hardest tier of FrontierMath (research-level problems).,,Numeric / symbolic answers checked automatically,% accuracy,,"Saturating at the very top (GPT-6.1 Sol 100%, GPT-6 Astra 97.6%, Claude Opus 5.5 95%) but wide spread below.","Very small number of problems, so each problem moves the score by several points.",https://epoch.ai/benchmarks
otis_mock_aime,OTIS Mock AIME 2024-2025,Epoch AI (problems from the OTIS mock AIME),2024,"Competition mathematics at the level of the American Invitational Mathematics Examination, on problems less likely to be in training data than the official AIME.",,Integer answers (0-999),% accuracy,,Saturated: several models at 100%.,Used here as the AIME proxy because Artificial Analysis retired AIME 2025 and no independent AIME 2026 table covering current models was found.,https://epoch.ai/benchmarks
arc_agi_2,ARC-AGI-2,ARC Prize Foundation,2025,Visual grid puzzles that are easy for humans but require learning a new rule from a few examples.,,"Grid output (pixel-exact), semi-private evaluation set",% of tasks solved,"Human panel 100% (ARC Prize data, about $17 per task).",Near saturation at the top (GPT-6 Astra 95%).,Scores depend on reasoning effort and cost; ARC Prize reports cost per task alongside score.,https://arcprize.org/arc-agi/2
arc_agi_3,ARC-AGI-3,ARC Prize Foundation,2026,"Interactive game environments where an agent must discover the goal and the rules by playing, scored on how efficiently it learns compared with humans.",,Interactive agentic games (no natural-language instructions),% score (100% = beats every game as efficiently as humans),100% by construction (all environments are human-solvable).,"Far from saturated with the standard harness (best 62.7%, GPT-6 Astra max effort).","ARC Prize also publishes a ""Provider Adapter"" harness where some models score near 100%; the two harnesses must not be mixed.",https://arcprize.org/arc-agi/3
swe_bench_verified,SWE-bench Verified,"SWE-bench (swebench.com), human-verified subset",2024,"Fixing real GitHub issues from open-source repositories, checked by running unit tests against the generated patch.",500,Code patches (agentic),% issues resolved,,"Saturated: 97.0% (Claude Opus 5, Vals.ai, mini-SWE-agent). Vals.ai no longer runs it on new releases and vendors stopped reporting it.",Strong contamination risk (public repositories); scores vary a lot with the agent harness; newest models (Sept 2026) not tested.,https://www.swebench.com/
swe_bench_pro,SWE-Bench Pro (public set),Scale AI,,"Harder, contamination-resistant software engineering tasks from copyleft and private codebases.","1,865 tasks in total; public set 731, private set 276, held-out 858",Code patches (agentic),% resolved,,"Not saturated, but the public SEAL leaderboard has not added models released after mid-2026 (latest entry Muse Spark 1.1).",Most current frontier models have no SEAL result; vendors use their own harnesses.,https://scale.com/leaderboard/swe_bench_pro_public
terminal_bench_4,Terminal-Bench 4.0,"Stanford University researchers, Laude Institute and open-source community",,"Completing hard multi-step tasks in a command-line terminal (software, system administration, data, ML, security).",66,"Agentic terminal tasks, each graded by its own test suite",% tasks passed (pass@1),,"Not saturated (best 63.6% independent, 70.6% vendor-reported).","Harness matters: AA uses mini-swe-agent (3 repeats), vendors use their own (Claude Code etc.); version 4.0 replaced 2.0/2.1 in 2026, so older Terminal-Bench numbers are not comparable.",https://github.com/harbor-framework/terminal-bench
tau3_banking,τ³-Banking (tau-bench family),Sierra,2026,A customer-support agent that must search a large knowledge base and make correct account changes while talking with a simulated user.,97,Dual-control agent-user simulation with tools,% pass@1 (5 repeats),,Not saturated (best about 51%).,Replaces τ²-Bench Telecom in AA reporting (τ²-Bench Telecom saturated above 95% for many models); removed from the AA Intelligence Index in v4.3 but still run as a standalone evaluation; not run for every new model.,https://github.com/sierra-research/tau2-bench
osworld_2,OSWorld 2.0 (binary completion),XLANG Lab (osworld-v2.xlang.ai),2026,"Long, realistic computer-use workflows on a real desktop (about 1.6 hours median for a human).",108 workflows,"Agentic computer use (GUI and tools), 500-step budget",% tasks fully completed,Median human operation time about 1.6 hours per task (no human success rate published).,Far from saturated (best 31.4% binary on release v2026.08.08).,"Several releases (v2026.06.24, v2026.08.08, v2.1) with different scores; vendors report OSWorld 2.1 or subsets with partial scores; very few models on the public board.",https://osworld-v2.xlang.ai/
osworld_2_partial,OSWorld 2.0 / 2.1 (partial score),XLANG Lab (osworld-v2.xlang.ai),2026,"Same tasks as OSWorld 2.0, with partial credit for sub-steps completed.",108 workflows,Agentic computer use,% partial score,,Not saturated.,"Mixes independent (2.0 leaderboard) and vendor-reported (2.1, offline subset) values: compare only within the same version and source.",https://osworld-v2.xlang.ai/
metr_time_horizon_50,METR 50% task-completion time horizon,METR,2025,"The length of software, ML and cybersecurity tasks (in human-expert time) that an AI agent completes with 50% probability.","Over a hundred tasks (RE-Bench, HCAST and shorter tasks), Time Horizon 1.1 suite",Agentic tasks; logistic fit of success vs. human task duration,minutes,Task lengths are calibrated on skilled professionals (about 5 years of experience on average).,Suite near its limit: METR states measurements above 16 hours are unreliable.,"No measurement published for any model released after April 2026 (last update 2026-05-08, Claude Mythos Preview at about 1,045 min); very wide confidence intervals.",https://metr.org/time-horizons
mmlu_pro,MMLU-Pro,TIGER-Lab,2024,"Broad academic knowledge and reasoning in 14 subjects, harder than the original MMLU.","Over 12,000 questions",Multiple choice (10 options),% accuracy,,Saturated: Vals.ai no longer runs it on new releases; top 92.4% (Claude Fable 5.1). Removed from the AA index in v4.0.,No independent results for models released after early September 2026.,https://arxiv.org/abs/2406.01574
livecodebench,LiveCodeBench,LiveCodeBench team (livecodebench.github.io),2024,"Competitive programming problems collected continuously from LeetCode, AtCoder and Codeforces to limit contamination.","Over 1,000 problems in v6 (May 2023-2025)",Code generation checked by hidden tests,% pass@1,,Saturated on easy and medium splits; Vals.ai no longer runs it on new releases; top 90.5%.,Results depend on problem date window and version; removed from the AA index in v4.0.,https://livecodebench.github.io/
simpleqa_verified,SimpleQA Verified,,2025,Short factual questions that test whether a model knows facts without inventing answers.,,Short open answers graded for correctness,% accuracy,,"Not saturated (best 75.6%, GPT-6 Astra).",Measures parametric knowledge without web search; models that abstain more score lower on accuracy but hallucinate less.,https://epoch.ai/benchmarks
aa_omniscience,AA-Omniscience,Artificial Analysis,,"Knowledge across six domains, rewarding correct answers and penalising confident wrong answers (hallucinations).","6,000",Open answers,index points (can be negative),,Not saturated (best 46.4).,A negative score means the model gives more wrong answers than right ones instead of abstaining; part of the AA Intelligence Index.,https://artificialanalysis.ai/methodology/intelligence-benchmarking
apex_agents,APEX-Agents,Mercor,2026,"Professional-services tasks completed by an agent working with files and tools, graded by rubric.",,"Agentic professional tasks, rubric-graded",% pass@1,,"Not saturated (best 82.2%, Gemini 4 Argon).",Run by the benchmark owner; values here are compiled by Epoch AI; standard errors of about 4-5 points.,https://epoch.ai/benchmarks

	id,name,creator,year,measures,items,format,unit,human_baseline,saturation,known_issues,url
	aa_intelligence_index,Artificial Analysis Intelligence Index v4.3.2,Artificial Analysis,2026,"One synthetic score that averages ten independent tests of agents, coding, scientific reasoning and general knowledge.","10 evaluations (AA-Briefcase, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, AA-LCR, AA-Omniscience, HLE, GDP.pdf, CritPt)","Composite index (weighted average: Agents 30%, Coding 20%, Scientific reasoning 20%, General 30%)",index points,,"Not saturated: top model 57.6 (Claude Opus 5.5, max effort) on 2026-10-06, unchanged on 2026-10-08.","Version changes often (v4.0 Jan 2026 to v4.3.2 Sep-Oct 2026); scores are not comparable across versions. Each reasoning-effort setting is scored separately. English, mostly text.",https://artificialanalysis.ai/methodology/intelligence-benchmarking
	epoch_capabilities_index,Epoch Capabilities Index (ECI),Epoch AI,,"One capability score fitted across dozens of benchmarks, so that models tested on different benchmarks can be compared on one scale.",,Statistical composite (item-response-style fit over many benchmarks),ECI points (with 90% CI),,Not saturated (open scale).,"Depends on which benchmarks each model has results for; very new models (e.g. Gemini 4 Argon, MiMo-V2.6-Pro, Mistral Large 4) not yet scored.",https://epoch.ai/benchmarks
	lmarena_text,LMArena Text Arena (overall),LMArena,,"Which answer people prefer in blind side-by-side chats, turned into a rating.","8,626,731 votes across 413 models (snapshot of 2026-10-02)",Pairwise human preference votes,Elo-style score,,"Not saturated, but the top 20 models sit within about 40 points with overlapping confidence intervals.","Measures preference, not correctness; sensitive to style and length (style control applied by default); new models have few votes and wide intervals (marked preliminary). The site now redirects from lmarena.ai to arena.ai ('formerly known as LMArena'); no new snapshot after 2026-10-02 as of 2026-10-08.",https://lmarena.ai/leaderboard/text
	gpqa_diamond,GPQA Diamond,"David Rein, Samuel R. Bowman et al. (GPQA paper)",2023,"Very hard graduate-level multiple-choice questions in biology, physics and chemistry that cannot be answered by simple web search.",198,Multiple choice (4 options),% accuracy,"PhD-level experts 65% (74% discounting clear mistakes); skilled non-experts with web access 34% (on the full GPQA set, per the paper abstract).","Near saturation: best scores around 95-96% (Epoch AI, Artificial Analysis). Removed from the AA Intelligence Index in v4.2.","Small set (198 questions), so differences of 1-2 points are noise; the paper notes that experts reach 74% once clear mistakes they identified are discounted, so a few answers are disputable and the practical ceiling is below 100%.",https://arxiv.org/abs/2311.12022
	hle,Humanity's Last Exam (HLE),Center for AI Safety and Scale AI,2025,"Expert-written closed-ended questions across dozens of academic subjects, designed to be the hardest academic exam for AI.","2,500 public questions (Artificial Analysis evaluates 2,158)","Exact-match short answers and multiple choice (24% MC, 14% need an image)",% accuracy,,"Not saturated: about 61% without tools (Claude Opus 5.5, AA); vendors report up to 67.7% with tools.","With-tools and no-tools scores are mixed in announcements; answer grading uses an LLM judge (judge choice differs between AA, Scale and vendors); CAIS released an updated ""HLE-Rolling"" set on 2026-09-17.",https://lastexam.ai/
	frontiermath_t1_3,"FrontierMath Tiers 1-3 (v2, private)",Epoch AI,2026,Unpublished research-level mathematics problems with automatically checkable answers.,,Numeric / symbolic answers checked automatically,% accuracy,,Approaching saturation for the very top (93.7% for GPT-6 Astra and GPT-6.1 Sol).,"Held privately and run only by Epoch AI, so results cannot be reproduced by third parties.",https://epoch.ai/benchmarks
	frontiermath_t4,"FrontierMath Tier 4 (v2, private)",Epoch AI,2026,The hardest tier of FrontierMath (research-level problems).,,Numeric / symbolic answers checked automatically,% accuracy,,"Saturating at the very top (GPT-6.1 Sol 100%, GPT-6 Astra 97.6%, Claude Opus 5.5 95%) but wide spread below.","Very small number of problems, so each problem moves the score by several points.",https://epoch.ai/benchmarks
	otis_mock_aime,OTIS Mock AIME 2024-2025,Epoch AI (problems from the OTIS mock AIME),2024,"Competition mathematics at the level of the American Invitational Mathematics Examination, on problems less likely to be in training data than the official AIME.",,Integer answers (0-999),% accuracy,,Saturated: several models at 100%.,Used here as the AIME proxy because Artificial Analysis retired AIME 2025 and no independent AIME 2026 table covering current models was found.,https://epoch.ai/benchmarks
	arc_agi_2,ARC-AGI-2,ARC Prize Foundation,2025,Visual grid puzzles that are easy for humans but require learning a new rule from a few examples.,,"Grid output (pixel-exact), semi-private evaluation set",% of tasks solved,"Human panel 100% (ARC Prize data, about $17 per task).",Near saturation at the top (GPT-6 Astra 95%).,Scores depend on reasoning effort and cost; ARC Prize reports cost per task alongside score.,https://arcprize.org/arc-agi/2
	arc_agi_3,ARC-AGI-3,ARC Prize Foundation,2026,"Interactive game environments where an agent must discover the goal and the rules by playing, scored on how efficiently it learns compared with humans.",,Interactive agentic games (no natural-language instructions),% score (100% = beats every game as efficiently as humans),100% by construction (all environments are human-solvable).,"Far from saturated with the standard harness (best 62.7%, GPT-6 Astra max effort).","ARC Prize also publishes a ""Provider Adapter"" harness where some models score near 100%; the two harnesses must not be mixed.",https://arcprize.org/arc-agi/3
	swe_bench_verified,SWE-bench Verified,"SWE-bench (swebench.com), human-verified subset",2024,"Fixing real GitHub issues from open-source repositories, checked by running unit tests against the generated patch.",500,Code patches (agentic),% issues resolved,,"Saturated: 97.0% (Claude Opus 5, Vals.ai, mini-SWE-agent). Vals.ai no longer runs it on new releases and vendors stopped reporting it.",Strong contamination risk (public repositories); scores vary a lot with the agent harness; newest models (Sept 2026) not tested.,https://www.swebench.com/
	swe_bench_pro,SWE-Bench Pro (public set),Scale AI,,"Harder, contamination-resistant software engineering tasks from copyleft and private codebases.","1,865 tasks in total; public set 731, private set 276, held-out 858",Code patches (agentic),% resolved,,"Not saturated, but the public SEAL leaderboard has not added models released after mid-2026 (latest entry Muse Spark 1.1).",Most current frontier models have no SEAL result; vendors use their own harnesses.,https://scale.com/leaderboard/swe_bench_pro_public
	terminal_bench_4,Terminal-Bench 4.0,"Stanford University researchers, Laude Institute and open-source community",,"Completing hard multi-step tasks in a command-line terminal (software, system administration, data, ML, security).",66,"Agentic terminal tasks, each graded by its own test suite",% tasks passed (pass@1),,"Not saturated (best 63.6% independent, 70.6% vendor-reported).","Harness matters: AA uses mini-swe-agent (3 repeats), vendors use their own (Claude Code etc.); version 4.0 replaced 2.0/2.1 in 2026, so older Terminal-Bench numbers are not comparable.",https://github.com/harbor-framework/terminal-bench
	tau3_banking,τ³-Banking (tau-bench family),Sierra,2026,A customer-support agent that must search a large knowledge base and make correct account changes while talking with a simulated user.,97,Dual-control agent-user simulation with tools,% pass@1 (5 repeats),,Not saturated (best about 51%).,Replaces τ²-Bench Telecom in AA reporting (τ²-Bench Telecom saturated above 95% for many models); removed from the AA Intelligence Index in v4.3 but still run as a standalone evaluation; not run for every new model.,https://github.com/sierra-research/tau2-bench
	osworld_2,OSWorld 2.0 (binary completion),XLANG Lab (osworld-v2.xlang.ai),2026,"Long, realistic computer-use workflows on a real desktop (about 1.6 hours median for a human).",108 workflows,"Agentic computer use (GUI and tools), 500-step budget",% tasks fully completed,Median human operation time about 1.6 hours per task (no human success rate published).,Far from saturated (best 31.4% binary on release v2026.08.08).,"Several releases (v2026.06.24, v2026.08.08, v2.1) with different scores; vendors report OSWorld 2.1 or subsets with partial scores; very few models on the public board.",https://osworld-v2.xlang.ai/
	osworld_2_partial,OSWorld 2.0 / 2.1 (partial score),XLANG Lab (osworld-v2.xlang.ai),2026,"Same tasks as OSWorld 2.0, with partial credit for sub-steps completed.",108 workflows,Agentic computer use,% partial score,,Not saturated.,"Mixes independent (2.0 leaderboard) and vendor-reported (2.1, offline subset) values: compare only within the same version and source.",https://osworld-v2.xlang.ai/
	metr_time_horizon_50,METR 50% task-completion time horizon,METR,2025,"The length of software, ML and cybersecurity tasks (in human-expert time) that an AI agent completes with 50% probability.","Over a hundred tasks (RE-Bench, HCAST and shorter tasks), Time Horizon 1.1 suite",Agentic tasks; logistic fit of success vs. human task duration,minutes,Task lengths are calibrated on skilled professionals (about 5 years of experience on average).,Suite near its limit: METR states measurements above 16 hours are unreliable.,"No measurement published for any model released after April 2026 (last update 2026-05-08, Claude Mythos Preview at about 1,045 min); very wide confidence intervals.",https://metr.org/time-horizons
	mmlu_pro,MMLU-Pro,TIGER-Lab,2024,"Broad academic knowledge and reasoning in 14 subjects, harder than the original MMLU.","Over 12,000 questions",Multiple choice (10 options),% accuracy,,Saturated: Vals.ai no longer runs it on new releases; top 92.4% (Claude Fable 5.1). Removed from the AA index in v4.0.,No independent results for models released after early September 2026.,https://arxiv.org/abs/2406.01574
	livecodebench,LiveCodeBench,LiveCodeBench team (livecodebench.github.io),2024,"Competitive programming problems collected continuously from LeetCode, AtCoder and Codeforces to limit contamination.","Over 1,000 problems in v6 (May 2023-2025)",Code generation checked by hidden tests,% pass@1,,Saturated on easy and medium splits; Vals.ai no longer runs it on new releases; top 90.5%.,Results depend on problem date window and version; removed from the AA index in v4.0.,https://livecodebench.github.io/
	simpleqa_verified,SimpleQA Verified,,2025,Short factual questions that test whether a model knows facts without inventing answers.,,Short open answers graded for correctness,% accuracy,,"Not saturated (best 75.6%, GPT-6 Astra).",Measures parametric knowledge without web search; models that abstain more score lower on accuracy but hallucinate less.,https://epoch.ai/benchmarks
	aa_omniscience,AA-Omniscience,Artificial Analysis,,"Knowledge across six domains, rewarding correct answers and penalising confident wrong answers (hallucinations).","6,000",Open answers,index points (can be negative),,Not saturated (best 46.4).,A negative score means the model gives more wrong answers than right ones instead of abstaining; part of the AA Intelligence Index.,https://artificialanalysis.ai/methodology/intelligence-benchmarking
	apex_agents,APEX-Agents,Mercor,2026,"Professional-services tasks completed by an agent working with files and tools, graded by rubric.",,"Agentic professional tasks, rubric-graded",% pass@1,,"Not saturated (best 82.2%, Gemini 4 Argon).",Run by the benchmark owner; values here are compiled by Epoch AI; standard errors of about 4-5 points.,https://epoch.ai/benchmarks