AI models, their providers' prices and rankings from independent measurements. Under every table it says where the figure comes from, who measured it, when, and how far it can be trusted.
Last recalculated: 7 August 2026
Overall score
The overall ranking: five abilities folded into one score so that models can be compared as a whole rather than on a single task. It is built from blind comparisons by real people, so it answers whose replies get preferred more often, not how many tasks were solved. If you need one specific skill, the dedicated collection is the more precise place to look.
Task score
How many programming tasks the model actually solved on fixed benchmark sets — a verifiable result, not an opinion. The collection has a second view, based on blind human ratings: the order there is different, and the gap between the two lists says more than either list on its own.
Task score
Reasoning tasks where the answer is either right or wrong: what counts is the share of solved problems, not of pleasing answers. Next to it there is a second view, from blind human ratings. Logic is where preference and a solved task part ways the most: a well-worded chain of reasoning pleases people even when its answer is wrong.
Task score
A test of factual knowledge on questions that have a single correct answer. It differs from the neighbouring collections in that the model does not reason or write code here — it recalls. The second view on the same topic comes from blind human ratings.
Task score
The share of solved mathematical problems, from olympiad to research level. The percentages are lower than in other collections, and that is a property of the problems, not of the models: comparing them with the numbers from the knowledge collection is meaningless. The second view comes from blind human ratings.
Task score
Here the model does not answer a question but does the work: edits files in an operating system, runs searches, completes tasks in a working environment. That is why the percentages are low even for strong models, and they cannot be compared with scores from the knowledge collection — it is different work, not different strength.
Arena Score, English
How models answer specifically in English: the score is built from blind comparisons in which people picked the better English reply. This is a collection about the language, not about ability — a model strong in the overall ranking can lose here to one that handles English phrasing better.
Arena Score, French
Blind comparison of replies in French. It needs a collection of its own because the order of models changes noticeably from language to language — and the overall ranking does not show that difference.
Arena Score, Spanish
Blind comparison of replies in Spanish. As with the other language slices, the interesting part is not the score itself but where a model stands relative to its own position in other languages: consistency across languages says more about a model than one high line.
Price, $ per 1M tokens
The price of API access per million tokens, in dollars, by the cheapest offer across providers. Input and output are folded into one number: output is usually several times more expensive, and ranking models by the input price alone would put something other than the cheapest on top.
Context, tokens
How much text the model accepts in a single request. The number is declared by the developer and is not something we can verify yet: on long windows, quality usually degrades well before the limit. So this collection answers whether the text will fit — not whether the model will make sense of it.
Arena Score, images
How well the model understands an image together with text: blind comparison of replies to prompts that come with a picture attached. Not to be confused with image generation — there the model draws, here it looks.
Arena Score, generation
Blind comparison of finished images: people pick the better picture without knowing which model drew it. This judges the result by eye rather than checking it against a task, which is why the difference between neighbouring places is most often statistically insignificant.
Arena Score, video
The same blind comparison as for images, but for video. There are an order of magnitude fewer models here than in the text collections, and the top places stand apart from the rest more visibly — simply because there are few participants.
Arena Score, web
Prompts about layout and web applications, rated blindly by people. It differs from the programming collection in what is measured: there, solved tasks are counted; here, whose result people preferred. For the web that is closer to the point, because how the result looks is part of what gets judged.
Arena Score, writing
Fiction and free-form writing, rated blindly by people: work of this kind has no correct answer, so preference is the most honest measure available. There is also a second view, from a scored run of set tasks; it measures something else and yields a different order.
Overall score
The same overall score as in the main ranking, but only among models with open weights — the ones you can download and run on your own hardware. The collection shows how far open models trail the closed ones, and in which tasks they do not trail at all.
The catalogue, the prices and the scores are collected from open sources; the right to republish is checked against the licence before a figure reaches the site. Under every table it says who measured what it shows and when. Figures we are not allowed to publish are used only for cross-checking and are not shown on the site.