The AI model leaderboard
Benchmarks
Per-benchmark AI leaderboards — GPQA, MMLU-Pro, SWE-Bench, HumanEval, LiveBench, Humanity's Last Exam, MATH and LMArena Elo. Every benchmark is normalized to one 0–100 scale and ranked from the public results published by each benchmark's maintainers.
Hugging Face
GPQA
GPQA is a multiple-choice benchmark built by Google DeepMind with 448 graduate-level science questions across physics, chemistry and biology. Its name means the questions are hard enough that search engines cannot answer them, which keeps the benchmark resistant to contamination. Scores reflect how often a model answers these expert questions correctly.
18 models ranked · #1 GPT-5 99 →
Hugging Face
MMLU-Pro
MMLU-Pro is the harder successor to MMLU, the classic knowledge benchmark. It expands the original test to over 12,000 questions across 14 academic disciplines and adds ten-option answers, making it far more difficult to guess. It measures the breadth of a model’s world knowledge and how reliably it reasons across topics from law and medicine to physics and economics.
18 models ranked · #1 GPT-5 98 →
BenchLM
HumanEval
HumanEval is OpenAI’s benchmark for code generation: 164 hand-written programming problems that test whether a model can produce a correct Python function from a docstring. Scores are reported as pass@1, the share of problems solved on the first attempt. It is one of the oldest and most widely cited coding benchmarks, even if real-world programming quickly outgrew it.
18 models ranked · #1 GPT-5 100 →
BenchLM
SWE-Bench
SWE-Bench is the agentic coding benchmark: it gives a model a real GitHub issue from an open-source repository and asks it to produce a patch that passes the project’s tests. With more than 2,000 tasks, it measures whether an AI can navigate unfamiliar codebases, understand bug reports and land fixes that actually merge. It is the closest public proxy for a hired software engineer.
18 models ranked · #1 GPT-5 93 →
LiveBench
LiveBench
LiveBench is a contamination-resistant benchmark released continuously by Abacus AI, with new question sets published on a fixed schedule so models cannot train on leaked answers. It scores 23 objective tasks across seven categories — reasoning, coding, mathematics, data analysis, language and more — where every answer is verifiable by code, not by an AI judge.
18 models ranked · #1 GPT-5 99 →
BenchLM
Humanity's Last Exam
Humanity's Last Exam is the hardest knowledge benchmark ever assembled: 2,500 questions written by more than 1,000 experts across 100 subjects, from mathematics and physics to law and linguistics. Designed by the Center for AI Safety and Scale AI as a frontier test, most top models still score below 20 percent. It is the best available measure of the ceiling of AI knowledge.
18 models ranked · #1 GPT-5 89 →
Hugging Face
MATH
The MATH benchmark is a collection of 5,000 competition-level mathematics problems in 7 subjects, from algebra and counting to number theory and precalculus. It tests not just calculation but step-by-step reasoning, since problems are graded on the final answer alone. It remains the standard signal for how well a model actually thinks — rather than how well it memorized.
18 models ranked · #1 GPT-5 95 →
LMArena
LMArena Elo
LMArena (formerly LMSYS Chatbot Arena) is the largest human-preference leaderboard in AI. Blind side-by-side battles between models are voted on by millions of real users, and the results feed a Bradley-Terry Elo rating — the same system chess players know. It captures what people actually prefer to talk to, which no automated benchmark can fully replicate.
18 models ranked · #1 GPT-5 100 →
AIRank aggregates these results from the public sources listed on each page. We do not run our own evaluations, and we normalize raw results to a 0–100 scale so tables are comparable.