LLM Benchmarks

Discover and compare AI model performance across 39 benchmarks.

Data from Artificial Analysis and LayerLens

8 Ways to Use Fewer Tokens

39 benchmarks

Terminal-Bench

Agentic · Coding agents

AI coding agents (Claude Code, Codex, Gemini CLI, …) ranked by pass rate on real end-to-end terminal tasks — scored per agent × model.

View agent leaderboard →

Humanity's Last Exam

Reasoning and Logic

Humanity's Last Exam — extremely challenging questions designed to test the upper limits of AI capability across diverse domains.

1
Anthropic
Claude Fable 5.1
59.1
2
Google
Gemini 4 Argon
57.1
3
Anthropic
Claude Fable 5
55.5
4
Anthropic
Claude Sonnet 5.5
55.0

GPQA

Reasoning and Logic

Graduate-level multiple-choice questions written by domain experts in biology, physics, and chemistry. Questions are Google-proof and extremely difficult.

1
OpenAI
GPT-6 Astra
96.1
2
Google
Gemini 3.8 Flash
95.3
3
SpaceXAI
Grok 4.6
94.9
4
Google
Gemini 3.7 Flash
94.5

LCR

Reasoning and Logic

Long Context Retrieval benchmark testing ability to find and use information in very long documents.

1
K
Kimi K3
88.7
2
X
MiMo v2.6 Pro
86.3
3
K
Kimi K3.1 Pro
86.3
4
Anthropic
Claude Fable 5.1
85.3

IFBench

Instruction Following

Instruction Following Benchmark measuring LLM ability to adhere to nuanced writing constraints and formatting requirements.

1
MiniMax
MiniMax M3
82.9
2
NVIDIA
Nemotron 3 Ultra 550B A55B
81.4
3
A
Qwen3.7 Max
80.5
4
X
MiMo v2.5 Pro
79.9

Tau2

Multi-turn

Tau2 benchmark testing multi-turn agent capabilities in airline and retail domains.

1
Z
GLM-5.2
99.1
2
Z
GLM-4.7-Flash
98.8
3
Z
GLM-5 Turbo
98.5
4
Z
GLM-5V Turbo
98.5

TerminalBench

Multi-turn

Terminal-based benchmark testing AI ability to interact with command-line interfaces and solve system tasks.

1
OpenAI
GPT-5.6 Sol
65.9
2
Anthropic
Claude Fable 5
62.9
3
OpenAI
GPT-5.5
60.6
4
Anthropic
Claude Opus 4.8
58.3

MMLU-Pro

General Knowledge

Massive Multitask Language Understanding benchmark testing knowledge across 57 diverse subjects including STEM, humanities, social sciences, and professional domains.

1
Google
Gemini 3 Pro Preview
89.8
2
Anthropic
Claude Opus 4.5
89.5
3
Google
Gemini 3 Flash Preview
89.0
4
Anthropic
Claude Opus 4.5
88.9

LiveCodeBench

Computer Science and Programming

Real-world coding benchmark with problems from competitive programming contests, testing code generation and problem-solving abilities.

1
Google
Gemini 3 Pro Preview
91.7
2
Google
Gemini 3 Flash Preview
90.8
3
DeepSeek
DeepSeek V3.2 Speciale
89.6
4
OpenAI
GPT-5.2
89.4

AIME 2025

Mathematical Problem Solving

American Invitational Mathematics Examination 2025 problems testing olympiad-level mathematical reasoning.

1
OpenAI
GPT-5.2 Pro
99.0
2
OpenAI
GPT-5 Codex
98.7
3
Google
Gemini 3 Flash Preview
97.0
4
DeepSeek
DeepSeek V3.2 Speciale
96.7

AGIEval English

Reasoning and Logic

AGIEval English — human-level reasoning tasks from standardized exams like SAT, LSAT, and civil service exams.

1
Google
Gemini 3.1 Pro Preview
94.0
2
Anthropic
Claude Opus 5
93.5
3
Google
Gemini 3.7 Flash
93.4
4
Google
Gemini 3.5 Flash
93.0

BBH

Reasoning and Logic

Big-Bench Hard — challenging subset of BIG-Bench focusing on tasks where language models previously underperformed.

1
Google
Gemini 3.1 Pro Preview
96.1
2
A
Qwen3.7 Max
95.8
3
A
Qwen3.8 Max
95.5
4
Qwen
Qwen3.6 Max Preview
95.2

MATH-500

Mathematical Problem Solving

Competition mathematics problems requiring multi-step reasoning, covering algebra, geometry, number theory, and calculus.

1
OpenAI
GPT-5
99.4
2
OpenAI
o3
99.2
3
SpaceXAI
Grok 3 Mini
99.2
4
Anthropic
Claude Sonnet 4
99.1

AIME 2024

Mathematical Problem Solving

American Invitational Mathematics Examination 2024 problems testing olympiad-level mathematical reasoning.

1
OpenAI
GPT-5
95.7
2
SpaceXAI
Grok 4
94.3
3
OpenAI
o4 Mini
94.0
4
A
Qwen3 235B A22B Thinking 2507
94.0

SciCode

Computer Science and Programming

Science coding benchmark measuring AI ability to solve scientific computing tasks.

1
Anthropic
Claude Fable 5.1
63.1
2
Google
Gemini 4 Argon
61.8
3
Anthropic
Claude Fable 5
61.0
4
Anthropic
Claude Sonnet 5.5
61.0

MMMU

Multimodal

Multimodal Understanding benchmark testing vision-language models on expert-level tasks.

1
Google
Gemini 3.1 Pro Preview
83.8
2
OpenAI
GPT-5.5
83.0
3
Google
Gemini 3.5 Flash
83.0
4
Google
Gemini 3.6 Flash
82.9

MMLU

General Knowledge

Massive Multitask Language Understanding — tests knowledge across 57 subjects.

1
A
Qwen3.7 Max
93.7
2
OpenAI
GPT-5
93.5
3
OpenAI
o3
93.1
4
A
Qwen3.5 397B A17B
92.7

SWE-bench Lite

Multi-turn

Software Engineering benchmark testing ability to resolve real GitHub issues.

1
Anthropic
Claude Opus 4.6
62.7
2
Anthropic
Claude Opus 4.6
62.7
3
MiniMax
MiniMax M2.5
56.3
4
OpenAI
GPT-5
54.3

HumanEval

Computer Science and Programming

OpenAI HumanEval benchmark measuring Python code generation from function docstrings.

1
Anthropic
Claude Sonnet 4.5
97.6
2
Anthropic
Claude Sonnet 4.5
97.0
3
Anthropic
Claude Opus 4.5
97.0
4
Anthropic
Claude Opus 4.5
97.0

MBPP Plus

Computer Science and Programming

Mostly Basic Python Problems Plus — tests Python code generation with enhanced test cases.

1
K
Kimi K2.6
66.9
2
A
Qwen3 235B A22B
66.2
3
A
Qwen3 235B A22B
66.2
4
A
Qwen3 32B
63.4

Accounting Audit

Financial Reasoning

Accounting and audit benchmark testing financial reasoning capabilities.

1
OpenAI
GPT-4.1
83.3
2
Google
Gemini 2.5 Flash
83.3
3
Google
Gemini 2.5 Flash
83.3
4
OpenAI
GPT-5
83.3

SimpleQA

Reasoning and Logic

Simple question answering benchmark testing factual accuracy and knowledge retrieval.

1
Google
Gemini 2.5 Pro
53.0
2
A
Qwen3 235B A22B Instruct 2507
50.6
3
A
Qwen3 VL 235B A22B Instruct
46.7
4
OpenAI
GPT-4.1
40.4

GSM8K

Mathematical Problem Solving

Grade School Math 8K — 8,500 high quality grade school math word problems.

1
Google
Gemini 2.5 Pro
95.7
2
NVIDIA
Nemotron 3 Nano 30B A3B
95.4
3
NVIDIA
Nemotron 3 Nano 30B A3B
95.4
4
DeepSeek
R1 0528
95.3

BBEH

Reasoning and Logic

Big Bench Extra Hard — even more challenging reasoning tasks pushing the limits of language model capabilities.

1
DeepSeek
DeepSeek V4 Flash (Non-Reasoning)
67.3
2
DeepSeek
DeepSeek V4 Pro
64.9
3
OpenAI
GPT-5
64.1
4
OpenAI
GPT-5 Mini
54.9

ARC Challenge

Reasoning and Logic

AI2 Reasoning Challenge (Challenge set) — grade-school science questions requiring complex reasoning.

1
OpenAI
GPT-5
96.3
2
Z
GLM 5
96.0
3
Z
GLM 5
96.0
4
DeepSeek
DeepSeek V4 Flash (Non-Reasoning)
95.4

BIRD-CRITIC

Multi-turn

BIRD-CRITIC — multi-turn benchmark testing SQL generation and database interaction.

1
Anthropic
Claude Opus 4.7
36.3
2
Anthropic
Claude Opus 4.6
34.0
3
Anthropic
Claude Opus 4.6
34.0
4
K
Kimi K2.6
33.3

AGIEval Chinese

Reasoning and Logic

AGIEval Chinese — reasoning tasks from Chinese standardized exams (Gaokao, civil service).

1
DeepSeek
DeepSeek V3.2 Exp
90.1
2
A
Qwen3 235B A22B
89.4
3
A
Qwen3 235B A22B
89.4
4
Z
GLM 4.6
88.2

ARC Easy

Reasoning and Logic

AI2 Reasoning Challenge (Easy set) — grade-school science questions.

1
A
Qwen3 32B
99.1
2
A
Qwen3 32B
99.1
3
Mistral
Mistral Medium 3
99.1
4
OpenAI
GPT-4.1
99.0

BFCL v3

Multi-turn

Berkeley Function Calling Leaderboard v3 — testing function/tool calling accuracy.

1
Z
GLM 4.5
76.7
2
Anthropic
Claude Opus 4.7
76.6
3
Google
Gemini 3.1 Flash Lite Preview
76.5
4
A
Qwen3 32B
75.7

IFEval

Instruction Following

Instruction Following Evaluation benchmark testing how well LLMs follow detailed formatting and content constraints.

1
K
Kimi K2.5
92.6
2
K
Kimi K2.5
92.6
3
Google
Gemini 2.5 Pro
90.8
4
Z
GLM 4.7
90.8

MedQA

Reasoning and Logic

Medical question answering benchmark from USMLE-style questions.

1
Google
Gemini 2.5 Pro
94.6
2
DeepSeek
R1
92.1
3
OpenAI
GPT-4.1
89.7
4
A
Qwen3 Max
85.5

Mathematics

Mathematical Problem Solving

Mathematics benchmark covering algebra, geometry, number theory, and calculus problems.

1
Anthropic
Claude Opus 4.6
95.6
2
Anthropic
Claude Opus 4.6
95.6
3
Z
GLM 5
94.0
4
Z
GLM 5
94.0

Stock BCS

Financial Reasoning

Stock market benchmark testing financial analysis capabilities.

1
A
Qwen2.5 72B Instruct
100.0
2
OpenAI
GPT-4.1
91.7
3
A
Qwen3 235B A22B
83.3
4
A
Qwen3 235B A22B
83.3

Knights and Knaves

Reasoning and Logic

Logic puzzle benchmark based on knights (truth-tellers) and knaves (liars) puzzles.

1
DeepSeek
DeepSeek V4 Pro
100.0
2
DeepSeek
R1 0528
97.9
3
OpenAI
GPT-OSS-20b
94.0
4
OpenAI
GPT-4.1
77.1

Formal Logic Extended

Reasoning and Logic

Extended formal logic benchmark testing deductive and propositional reasoning.

1
DeepSeek
R1 0528
99.2
2
OpenAI
GPT-4.1
90.6
3
Deepseek
DeepSeek V3
83.6
4
Cohere
Command A
80.7

GAIA

Multi-turn

GAIA — General AI Assistants benchmark testing multi-step real-world tasks.

1
OpenAI
GPT-5 Mini
44.8
2
Google
Gemini 2.5 Pro
33.3
3
DeepSeek
R1 0528
27.9
4
Mistral
Mistral Medium 3.1
23.3

WMDP

Reasoning and Logic

Weapons of Mass Destruction Proxy — benchmark testing knowledge safety boundaries.

1
Google
Gemini 3 Flash Preview
86.8
2
Google
Gemini 3 Flash Preview
86.8
3
Deepseek
DeepSeek V3
71.8
4
OpenAI
GPT-4.1
71.4

DarkBench

Reasoning and Logic

DarkBench — benchmark testing model safety and resistance to adversarial attacks.

1
A
Qwen3 32B
49.2
2
A
Qwen3 32B
49.2
3
Google
Gemini 3 Flash Preview
49.1
4
Google
Gemini 3 Flash Preview
49.1

WMT 2014

Multilingual

Workshop on Machine Translation 2014 — multilingual translation quality benchmark.

1
Meta
Llama 4 Maverick
38.0
2
OpenAI
GPT-4.1
37.6
3
OpenAI
GPT-4o
37.5
4
Meta
Llama 4 Scout
37.1

MultiChallenge

General Knowledge

MultiChallenge — general knowledge benchmark with diverse challenge types.

1
Mistral
Mistral Medium 3.1
37.4
2
Amazon
Nova Pro 1.0
19.0