Рейтинг моделей ШІ

How this leaderboard works

Every figure comes from our catalog, our pricing rules, our own job records or, for quality, the public Arena leaderboard. What we cannot source, we leave out.

What is listed

Only models you can run right now: active and with a verified price. Left out: 130 catalog entries without a price (they cannot run), 0 removed models and 5 duplicate entries of the same provider model. Legacy models stay listed and link their successor.

Price

What one call with the model’s default settings costs on Railwail, from the same pricing rules that bill your runs (1 credit = $0.01). Token models show the price per 1M input and output tokens. Prices marked ≈ are billed by GPU time: the figure is a typical run, the charge follows the real run time.

Median run time

Measured on completed Railwail jobs: the median of the newest up to 500 runs in the last 90 days, shown from 5 runs. It includes queueing at the provider. For language models it depends on the answer length, so the LLM board is never sorted by speed. Measured today: 3 models.

Quality

Arena scores from the public Arena leaderboard (formerly LMArena, CC BY 4.0), where people compare two anonymous models on the same prompt and vote. A model gets a score only when the arena entry is exactly the model and default setting we run (reviewed list, 56 models today). No score is estimated.

Context window

The maximum tokens per request as listed in our catalog. Not re-measured by us.

Usage

Runs per model are only shown once they say something: at least 5 models with runs from 3 or more different accounts in 30 days. Until then there is no popularity ranking.

What we don’t show

No votes of our own and no estimated scores: our own arena has 0 decided battles and 2 model votes, far too few for a rating. Models the public Arena has not rated in the setting we run show “No Arena rating”.

Source:arena.ai leaderboardDatasetCC BY 4.0

Found a wrong price or a missing model? Write to us