How scoring worksHow scoring works

1. 01

   Every kind of model gets its own report card

   You wouldn’t grade a fish on climbing trees. Image models are graded on things like picture quality and following instructions; chat models on reasoning, accuracy, and speed. Each category has its own subjects.

2. 02

   Each subject gets a mark from 0 to 10

   Reasoning: 9. Accuracy: 8. Writing: 7. Every mark comes from a public measurement or a published checklist, never from a gut feeling.

3. 03

   Some subjects count more than others

   For a chat model, reasoning counts for 45% of the final grade and writing for 15%. The weight of every subject is shown next to it (w45, w15).

4. 04

   Price isn’t one of the subjects

   What a model costs doesn’t change what it can do, so price and availability are measured and shown on every card but kept out of the grade. Otherwise a cheap model would outrank a stronger one in a ranking that claims to say which is better.

5. 05

   A mark says how sure we are

   Where a source publishes a margin of error, we take the cautious end of it, and a model tested only a few hundred times doesn’t get a mark at all. Two models whose margins overlap are shown as tied, because the evidence can’t separate them.

6. 06

   The final score is a weighted average

   Strong marks in heavy subjects pull the score up more than strong marks in light ones. That’s the number in the corner of a card, like 8.2/10.

7. 07

   No final grade without enough homework

   A model needs marks covering at least 80% of what counts before it gets a final score, so one lucky subject can’t make it look amazing.

8. 08

   Nobody can buy a grade

   Companies don’t pay to be listed or scored, and they don’t see scores before they’re published. If a number is wrong, the fix is better evidence, not a phone call.

• ▸Each model is scored on one named version, shown on its page, together with the exact leaderboard entry behind each mark.
• ▸Cost uses a blended price: 3 parts input to 1 part output, per million tokens.
• ▸Rule B marks are absolute. A model’s mark changes only when its own measurement changes, never because another model was added to or removed from the directory.
• ▸A human-preference rating built on fewer than 500 votes is treated as unmeasured rather than as a weak result.
• ▸Where a model has several tested settings, we credit the strongest result we can be confident in: the highest lower bound, not the highest headline figure.
• ▸If there’s no trustworthy source for a criterion, it stays empty (shown as “—”). We never estimate or fill gaps.
• ▸Vendor marketing claims are never used as evidence. Independent measurements and published benchmark results are.

None of this makes a score the last word. Every leaderboard behind it measures what it can measure, on the prompts it happens to collect, at one moment in time. What this method adds is that each number is traceable to a dated, public source, computed by published rules, and honest about how firm it is.

1. v2.3Sep 26, 2026The two Arabic boards were merged into one table with a single combined figure, after being shown as separate tabs. Both boards’ own results stay visible on every row, because they disagree about some models and the disagreement is the finding. The combination fits a line between the two scales on the models both have run and weights each figure by how firm it is, so a converted result moves the answer least; the conversion’s error is carried into the band, which is why models known through one board alone are marked low and tie with much of the table.
2. v2.2Sep 26, 2026A multilingual board was added to the home page, built on LMArena’s non-English category: the same blind votes as the text arena, split by the language of the prompt. Each model is shown with its place on the English board beside it, and the site states both places without turning the difference between them into a claim — in a field of 400 models the ratings around any place sit inside one another’s confidence bands, so a move of even twenty places usually can’t be told apart from noise. The non-English arena’s scale is published with the others below. No score changed: this board sits beside the directory’s scores rather than feeding them.
3. v2.1Sep 26, 2026Arabic ability became measurable. Two independent Arabic leaderboards — AraGen, which has a judge model score free-form Arabic answers on six qualities, and the Open Arabic LLM Leaderboard, which averages seven native Arabic benchmarks — now publish results in a machine-readable form, so the site reads them directly and ranks the Arab world’s models on them. Marks follow Rule A, the low end of the published interval where there is one. No existing score changed: these boards are shown beside the directory, not folded into it, because they measure something different against a different field of models.
4. v2.0Sep 20, 2026Three changes, after the old method produced rankings that didn’t survive scrutiny. (1) Rule B no longer ranks a model against the others in the directory; it places the measurement on a fixed, published scale, so gaps reflect real differences, leading a category no longer means an automatic 10, and adding a model doesn’t move everyone else’s marks. (2) Where a source publishes a confidence interval, the mark uses the low end of it, a rating from fewer than 500 votes is treated as unmeasured, and every score carries a band; models whose bands overlap share a rank. (3) Price and availability are still measured and shown, but no longer count toward the score, which now measures capability only. Every score on the site changed.
5. v1.2Sep 19, 2026Coverage bar raised from 60% to 80%. A model now needs marks covering at least 80% of its rubric’s weight before it gets an overall score, so models with big gaps can’t outrank fully measured ones. Models below the bar keep their individual marks and show as in assessment.
6. v1.1Sep 19, 2026Rubrics revised before any scores were published so that every criterion has exactly one independent source. Criteria with no independent source yet (speed, image prompt adherence, video motion, and others) were removed until one exists. Audio, search, and enterprise tools stay unscored for the same reason. First scores published.
7. v1.0Sep 19, 2026Methodology published: rubrics, three marking rules, weighted average, 60% coverage bar.