40 families · 12 categories · 15,707 observations

Benchmarks

A family groups the versions of one benchmark (task sets change; scores across versions don’t mix). Categories with a composite weight feed the BenchAtlas Index; the rest are display-only.

Agentic Coding

Benchmark	Maintainer	Versions	Tasks	Observations
LiveBench Agentic Coding Contamination-limited monthly-refreshed benchmark; category mean over its tasks.	LiveBench	1	—	36
SWE-bench agent + model Resolving real GitHub issues from popular Python (and multilingual/multimodal) repositories. Results are agent+model systems — the scaffold matters.	SWE-bench	6	500	400
τ²-Bench Agentic tool-use benchmark (dual-control customer-service domains). Reported per model by evaluators running the standard τ² harness.	—	1	—	294
Terminal-Bench agent + model End-to-end terminal tasks executed by agent+model pairs.	Laude Institute / Harbor	4	—	496

Coding

Benchmark	Maintainer	Versions	Tasks	Observations
LiveBench Coding Contamination-limited monthly-refreshed benchmark; category mean over its tasks.	LiveBench	1	—	36
LiveCodeBench Competitive-programming problems with rolling contamination-free windows.	LiveCodeBench	2	1,055	190
SciCode Scientific-computing code generation from research problems.	—	1	—	235

Data Analysis

Benchmark	Maintainer	Versions	Tasks	Observations
LiveBench Data Analysis Contamination-limited monthly-refreshed benchmark; category mean over its tasks.	LiveBench	1	—	36

External Composite Indices

Benchmark	Maintainer	Versions	Tasks	Observations
AA Coding Index Artificial Analysis's coding composite. Displayed for reference only.	Artificial Analysis	1	—	94
AA Intelligence Index Artificial Analysis's composite intelligence index. Displayed for reference; excluded from the BenchAtlas Index (its inputs already feed it). Comparable only within a major index version.	Artificial Analysis	1	—	243
AA Math Index Artificial Analysis's math composite. Displayed for reference only.	Artificial Analysis	1	—	109
Epoch Capabilities Index Epoch AI's composite capabilities index (ECI). Displayed for reference only.	Epoch AI	1	—	444
Vals Index Vals AI's GDP-weighted composite over professional-domain benchmarks. Displayed for reference only.	Vals AI	1	—	35

Factuality

Benchmark	Maintainer	Versions	Tasks	Observations
MASK Honesty-under-pressure benchmark: does the model contradict its own beliefs?	Scale (Scale Labs)	1	—	67
SimpleQA Verified Short-form factuality benchmark (verified subset).	—	1	—	64

Human Preference

Benchmark	Maintainer	Versions	Tasks	Observations
LMArena Text Head-to-head human preference votes, Bradley–Terry Arena Score (Elo-like, style-controlled by default). Pool-relative: never comparable across arenas.	LMArena (Arena)	29	—	8,492

Knowledge & Science

Benchmark	Maintainer	Versions	Tasks	Observations
GPQA Diamond Graduate-level, Google-proof science questions (diamond subset).	—	1	198	537
Humanity's Last Exam Frontier-difficulty closed-ended academic questions (CAIS × Scale).	—	1	—	284
MMLU-Pro Harder, reasoning-focused successor to MMLU.	—	1	—	245

Language

Benchmark	Maintainer	Versions	Tasks	Observations
LiveBench Language Contamination-limited monthly-refreshed benchmark; category mean over its tasks.	LiveBench	1	—	36

Long Context & Instruction

Benchmark	Maintainer	Versions	Tasks	Observations
AA Long Context Reasoning Artificial Analysis's long-context reasoning benchmark.	Artificial Analysis	1	—	220
IFBench Instruction-following benchmark (verifiable constraints).	—	1	—	207
LiveBench Instruction Following Contamination-limited monthly-refreshed benchmark; category mean over its tasks.	LiveBench	1	—	36
MultiChallenge Multi-turn conversational instruction-following benchmark (Scale Labs).	Scale (Scale Labs)	1	—	30

Multimodal

Benchmark	Maintainer	Versions	Tasks	Observations
MMMU Massive multi-discipline multimodal understanding benchmark.	—	1	—	82
VISTA Visual-language understanding benchmark (Scale Labs).	Scale (Scale Labs)	1	—	63

Professional Domains

Benchmark	Maintainer	Versions	Tasks	Observations
CaseLaw Vals AI's case-law research and citation benchmark.	Vals AI	1	—	54
CorpFin Vals AI's corporate-finance document analysis benchmark.	Vals AI	1	—	122
LegalBench Legal reasoning task suite run by Vals AI.	Vals AI	1	—	124
MedQA Medical licensing-exam question answering.	Vals AI	1	—	95
TaxEval Vals AI's tax-law reasoning benchmark.	Vals AI	1	—	127

Reasoning & Math

Benchmark	Maintainer	Versions	Tasks	Observations
AIME American Invitational Mathematics Examination problems (recent years).	—	1	—	269
ARC-AGI Abstraction and reasoning over novel grid puzzles; semi-private eval split.	ARC Prize Foundation	3	—	1,226
EnigmaEval Long, multimodal puzzle-solving benchmark (Scale Labs).	Scale (Scale Labs)	1	—	47
FrontierMath Research-level mathematics problems maintained by Epoch AI.	Epoch AI	2	—	173
LiveBench Mathematics Contamination-limited monthly-refreshed benchmark; category mean over its tasks.	LiveBench	1	—	36
LiveBench Reasoning Contamination-limited monthly-refreshed benchmark; category mean over its tasks.	LiveBench	1	—	36
MATH-500 500-problem subset of the MATH competition dataset.	—	1	—	125
MATH Level 5 Hardest tier of the MATH competition dataset.	—	1	—	108
OTIS Mock AIME Mock AIME 2024–2025 problems evaluated by Epoch AI.	Epoch AI	1	—	154

Tasks = task count of the family’s current version, where the maintainer publishes one. Observations count every score we have ever captured for the family across all sources (append-only — scores that vanish upstream don’t un-happen).