Comparing AI Models
A leaderboard ranks AI models on a fixed set of tasks. It is a good first filter when choosing a model, provided you know what it measures, because each one measures something different: answers to exam-style questions, which reply people prefer in a blind comparison, whether a model can fix real software or call tools correctly, and what a model costs and how fast it answers.
Three limits apply to all of them. A high score on a public test can partly reflect test questions that found their way into the model's training data. Rankings move every few weeks as new models are released and retested, so check the date. And price and speed depend on the provider serving the model as well as on the model itself.
No leaderboard tells you whether a model, or a tool built on one, is fit for your organisation: what happens to the data you send, what the contract allows, which law applies and where it is hosted. That takes testing on your own prompts and data, and a reading of the terms.
Where to compare models
Artificial Analysis(opens in a new tab)
An independent analysis company
Its own intelligence indices, from evaluations it runs itself, plus output speed, latency and price for each API provider that serves a model.
When cost and speed matter as much as quality.
HELM(opens in a new tab)
Center for Research on Foundation Models, Stanford University
Holistic Evaluation of Language Models: many scenarios and several measures, not only accuracy, with the prompts and model outputs published so the results can be checked.
For a broad and transparent academic comparison.
SWE-bench(opens in a new tab)
Researchers at Princeton University
Whether a model or coding agent can resolve real issues taken from open-source GitHub projects, checked against each project’s own tests.
For coding assistants and software agents.
Berkeley Function Calling Leaderboard(opens in a new tab)
Gorilla project, UC Berkeley
How accurately a model calls functions and tools, including calls spread over several steps and turns of a conversation.
For agents that act through other systems.
Epoch AI Benchmarking Hub(opens in a new tab)
Epoch AI, an independent non-profit research institute
Results on demanding benchmarks, some run by Epoch itself and some collected from other sources.
For the hardest reasoning and mathematics benchmarks.
Arena (formerly LMArena, Chatbot Arena)(opens in a new tab)
Started at UC Berkeley in 2023, now a company
People compare two anonymous answers to the same prompt and vote for the better one. The votes give a ranking by human preference, for text, code, images and video.
To see which answers people prefer. Preference is not the same as correctness.
Links checked on 27 October 2026. We are not affiliated with any of these organisations and do not republish their results.
Choosing a model for a real use case?
Tell us about the use case and we will go through the options with you, including how each one handles your data and what its terms allow. No commitment, no cold-selling.
Common questions
Why look at more than one leaderboard?
Because each one measures something different. A model can lead on exam-style questions and trail on fixing real code or calling tools. Compare the leaderboards that match the work you need the model to do.
Can leaderboard results replace testing on our own data?
No. They narrow the field. Before committing, test the shortlisted models on your own prompts, data, latency targets and security constraints, and read the terms for how your data is handled.