Which AI is best today? With a source for every number.

It depends on what you mean by best. For capability measured on hard tests, the top of the ECI. For the preference of ordinary people in everyday conversations, the top of the ELO. LiveBench is a third reading, with refreshed questions to reduce the chance that the model has already seen the test. The rankings do not agree, which is why this page shows the columns side by side instead of inventing a single score.

Epoch Capabilities Index, from Epoch AI. A composite index that combines a model performance across several hard benchmarks into a single scale, which allows comparing models that never took exactly the same tests. The higher, the better.

It is the Arena human preference score. People send the same question to two models without knowing which they are, pick the better answer, and the system calculates a score on the same rating scheme as chess. It measures what people like to receive, which is not the same as raw capability.

It is a benchmark with 23 objective tasks in 7 categories (reasoning, coding, agentic coding, mathematics, data analysis, language and instruction following), with questions refreshed every six months to reduce the chance that the model has already seen the test. The score goes from 0 to 100 and is the average of the categories, the same formula the LiveBench website uses. It usually publishes a release's score before the Epoch AI index.

Because that model was not found in the Arena with certainty. Matching names by similarity is how a ranking starts lying without throwing any error. When there is no exact match, the column stays empty. We prefer the gap to the wrong number.

If it came out recently, it shows up in the release card at the top of the page and in the "Recent releases, not scored yet" list, right below the table. Epoch AI registers the model on release day and computes the capability score later, so there is a window when the model is public but has no ECI yet. In that window it does not take a position in the ranking, which is ordered by capability, but it appears with its LiveBench and Arena scores when they already exist, which usually happens before the ECI. We do not estimate the missing score.

They do, and you can filter for them alone in the bar above the table. The Access column shows whether the model is served through an API, whether the weights are open or whether access is restricted, following Epoch AI own classification.

The original data comes from the three sources credited below: Epoch AI and Arena under Creative Commons Attribution 4.0, and LiveBench under Attribution-ShareAlike 4.0. You can use it, commercially too, as long as you credit the original sources the same way we do here; anything taken from LiveBench stays under the same CC BY-SA. The credit is not a courtesy: it is the condition of the license.

Six times a day, automatically. The date and time of the last update are at the top of the page and in the table footer. If one of the sources is down, the page keeps the numbers from the previous update instead of blanking the column.