Rankings

Best right now.

Which AI models lead independent tests today, in plain English. The scores come from independent testers, never from the labs' own announcements, and refresh a few times a day.

Which AI should I use? →Compare two models →

Best all-rounder

Which model is the smartest overall?

GPT-6 Astra, GPT-6.1 Sol, Claude Opus 5.5 and GPT-6.1 Terra are effectively tied at the top of Epoch AI's combined test score: the gaps between them are within the margin of error.

What the score means: Epoch AI combines results from many hard tests of Reasoning model: A model that thinks step by step before answering. Slower, but much better at maths, code and tricky questions., maths, science, coding and knowledge into one score. It's set so that GPT-5 (2025) scores 150 and Claude 3.5 Sonnet (2024) scores 130, so 10 points is a big step. Models within about 3 points of each other are too close to call.

Epoch Capabilities Index · Epoch AI

People's favourite

Whose answers do people like best?

In blind votes, Gemini 4 Carbon and Gemini 4 Argon are neck and neck: Gemini 4 Carbon's answers win about 51% of the time.

What the score means: On LMArena, people ask a question, get answers from two anonymous models and vote for the better one. The votes become a rating, like a chess rating: 100 points ahead means winning about 64% of the time. This view adjusts for length and formatting, so long or flashy answers don't win on style alone.

LMArena text, adjusted for style · LMArena

Best for coding

Which model builds the best apps?

Claude Opus 5.5 builds the apps people like best: in blind votes it beats Claude Sonnet 5.5 about 56% of the time.

What the score means: On LMArena's WebDev arena, people ask two anonymous models to build the same web app and vote for the better result. The votes become a rating, like a chess rating: 100 points ahead means winning about 64% of the time.

LMArena WebDev · LMArena

Best free to download

What's the best model you can run yourself?

DeepSeek V4.1 Flash is the strongest model you can download and run yourself, scoring 159 against 169 for the best model overall.

What the score means: The same combined score, counting only models whose files anyone can download and run on their own computers or servers. Some come with licences that limit commercial use.

Epoch Capabilities Index, open-weight models · Epoch AI

Best image generator

Which model makes the best pictures?

GPT-Image-2.5 Sunburst makes the pictures people like best in blind votes, beating GPT-Image-2.6 Sunburst about 53% of the time.

What the score means: People describe a picture, see what two anonymous image models make of it, and vote for the better one. The votes become a rating, like a chess rating: 100 points ahead means winning about 64% of the time.

LMArena text-to-image · LMArena

Best video generator

Which model makes the best videos?

Gemini Omni 1.2 Flash and Gemini Omni 1.1 Flash make the videos people like best, neck and neck in blind votes.

What the score means: People describe a scene, watch what two anonymous video models make of it, and vote for the better clip. The votes become a rating, like a chess rating: 100 points ahead means winning about 64% of the time.

LMArena text-to-video · LMArena

Progress

How fast is AI improving?

A year ago the best model, GPT-5 Pro, scored 151 on Epoch AI's combined score. Today's best, GPT-6 Astra, scores 169: 18 points more in a year.

How to read it: each faint dot is a model Epoch AI has scored, placed at its release date. The line steps up each time a new model beats every earlier one. On this scale, 20 points is the gap between Claude 3.5 Sonnet (2024) and GPT-5 (2025).

How much to trust this

A Benchmark: A standard test used to compare models, like an exam. Useful, but labs tend to show the tests they do well on, so one score never tells the whole story. is a standard test, like an exam. It's the fairest way to compare models, but it isn't the whole story: tests cover some skills and miss others, labs can tune a model for well-known tests, and a few points apart is often too close to call.

So treat this as a guide to which models are at the front, not a precise ranking, and try the leaders on your own tasks. A model that is new may not be tested yet.

Data from

Epoch AI · Capabilities & benchmarking data, CC BY 4.0
LMArena · Leaderboard dataset, CC BY 4.0

Updated 5h ago