LLM benchmarks by field
Grouped by field: what each benchmark measures, its type and the number of models listed.
Coding
| Benchmark | What it measures | Type | Models |
|---|---|---|---|
| Coding | Rating on coding questions | Human ratings | 387 |
| LCB (English questions) | Recent competitive programming problems | Task results | 68 |
| MirrorCode | Reimplementing whole programs, long tasks | Task results | 10 |
| SWE-Bench verified | Fixing real Python bugs | Task results | 32 |
| Aider polyglot | Editing code in six programming languages | Reported by others | 55 |
| Terminal Bench | Work in a terminal | Reported by others | 44 |
Math
| Benchmark | What it measures | Type | Models |
|---|---|---|---|
| Math | Rating on math problems with definite answers | Human ratings | 379 |
| AIME (English questions) | Problems from a US math exam | Task results | 68 |
| FrontierMath 2025-02 | Unpublished math problems (old version) | Task results | 71 |
| FrontierMath Tier 4 2025-07 | Research-level math problems (old version) | Task results | 56 |
| FrontierMath Tier 4 v2 | Research-level math problems | Task results | 71 |
| FrontierMath Tiers 1-3 v2 | Advanced undergraduate to early research math | Task results | 88 |
| MATH level 5 | Math competition problems, top difficulty level | Task results | 97 |
| MATH500 (English questions) | A selection of math competition problems | Task results | 68 |
| OTIS Mock AIME 2024-2025 | AIME-style math competition problems | Task results | 201 |
| ProofBench | Math proofs | Reported by others | 81 |
Knowledge and reasoning
| Benchmark | What it measures | Type | Models |
|---|---|---|---|
| English hard prompts | Rating on hard English questions | Human ratings | 390 |
| Expert | Rating on expert-level questions | Human ratings | 343 |
| Hard prompts | Rating on hard questions | Human ratings | 392 |
| GPQA (English questions) | Expert-level biology, chemistry, physics questions (English questions in a Japanese evaluation suite) | Task results | 68 |
| GPQA diamond | Expert-level biology, chemistry, physics questions (the subset only experts answered correctly) | Task results | 221 |
| HellaSwag (English questions) | Choosing how an everyday scene continues | Task results | 68 |
| MMLU-Pro (English questions) | Hard multiple-choice questions across many fields | Task results | 68 |
| SimpleQA Verified | Answering short factual questions correctly | Task results | 83 |
| ARC-AGI-2 | Abstract grid puzzles | Reported by others | 91 |
| HLE | The hardest academic questions | Reported by others | 45 |
Writing
| Benchmark | What it measures | Type | Models |
|---|---|---|---|
| Creative writing | Rating on creative writing requests | Human ratings | 390 |
| Instruction following | Rating on following detailed instructions | Human ratings | 392 |
| Longer query | Rating on long questions | Human ratings | 370 |
| Multi-turn | Rating on multi-turn conversations | Human ratings | 390 |
| Lech Mazur Writing | Grading short stories | Reported by others | 43 |
Agents and documents
| Benchmark | What it measures | Type | Models |
|---|---|---|---|
| APEX-Agents | Professional work tasks | Reported by others | 51 |
| DeepResearch Bench | Answering by searching the web | Reported by others | 25 |
| Fiction.LiveBench | Reading long stories | Reported by others | 39 |
| GDP.pdf | Reading PDFs | Reported by others | 39 |
| OSWorld 2.0 | Operating a computer screen | Reported by others | 9 |
| The Agent Company | Simulated company work | Reported by others | 14 |
Japanese and other languages
| Benchmark | What it measures | Type | Models |
|---|---|---|---|
| Chinese | Rating on Chinese questions | Human ratings | 369 |
| English | Rating on English questions | Human ratings | 392 |
| French | Rating on French questions | Human ratings | 279 |
| German | Rating on German questions | Human ratings | 291 |
| Japanese | Rating on Japanese questions | Human ratings | 262 |
| Korean | Rating on Korean questions | Human ratings | 261 |
| Non-English | Rating on non-English questions | Human ratings | 392 |
| Polish | Rating on Polish questions | Human ratings | 213 |
| Russian | Rating on Russian questions | Human ratings | 356 |
| Spanish | Rating on Spanish questions | Human ratings | 279 |
| GPQA (Japanese version) | Expert-level biology, chemistry, physics questions (Japanese version) | Task results | 68 |
| JamC-QA | Multiple-choice questions on Japanese culture and daily life | Task results | 68 |
| JHumanEval | Writing Python code from Japanese descriptions | Task results | 68 |
| MIFEvalJa | Following format instructions given in Japanese | Task results | 68 |
| MIFEvalJa (measured with the English tasks) | Following format instructions given in Japanese | Task results | 68 |
| MMLU-ProX (Japanese version) | Hard multiple-choice questions across many fields (Japanese version) | Task results | 68 |
| PolyMathJaHT | Hard math problems (in Japanese) | Task results | 68 |
| WMT20enja | Translation from English to Japanese | Task results | 68 |
| WMT20jaen | Translation from Japanese to English | Task results | 68 |
Industry
| Benchmark | What it measures | Type | Models |
|---|---|---|---|
| Business and finance | Rating on business and finance questions | Human ratings | 385 |
| Entertainment and media | Rating on entertainment and media questions | Human ratings | 390 |
| Legal and government | Rating on legal and government questions | Human ratings | 365 |
| Medicine and healthcare | Rating on medicine and healthcare questions | Human ratings | 363 |
| Quantitative work | Rating on quantitative work questions | Human ratings | 376 |
| Science | Rating on science questions | Human ratings | 390 |
| Software and IT | Rating on software and IT questions | Human ratings | 392 |
| Writing and language work | Rating on writing and language work questions | Human ratings | 391 |
Other
| Benchmark | What it measures | Type | Models |
|---|---|---|---|
| Overall | Overall rating from all votes | Human ratings | 392 |
| Overall (excluding ties) | Rating without tied votes | Human ratings | 392 |
| Chess Puzzles | Choosing the best chess move | Task results | 85 |
| EBR-bench | Improving through repeated play | Task results | 27 |
| Furniture Assembly | Spotting furniture assembly mistakes | Task results | 34 |
| Mystery Game Puzzles | Best moves in an undisclosed game | Task results | 79 |
| SpireClimb | Winning a card game at its hardest setting | Task results | 7 |
How to read this page
- "Task results" are runs made by the evaluating organization itself under the same conditions; "Reported by others" collects figures published by the benchmark makers and others. Within each group, benchmarks are listed by type, then by name.
Data: Arena, Epoch AI, Swallow LLM team (CC BY 4.0), Models.dev (MIT), updated 2026-10-24, data sources