Where our scores come from
As of November 11, 2026, model pages show 6,349 benchmark results for 592 models. The model’s developer published 54% of them, the benchmark’s own leaderboard 37%, and independent evaluators and other third parties the other 9%. Every result links to its source, and a developer’s result never counts as independent evidence for a rank.
54% of the results come from the model’s developer
The table counts each result on a model page once, under whoever published it. The three columns are every model page, the models that hold a rank, and the 60 highest-ranked.
| Published by | Label on the row | Every model page6,349 results, 592 models | Ranked models4,193 results, 193 models | Top 60 by rank2,376 results, 60 models |
|---|---|---|---|---|
| The model’s developerLabeled Provider exact | Provider exact | 3,406 (54%) | 2,256 (54%) | 1,162 (49%) |
| The benchmark’s own leaderboardLabeled Benchmark exact | Benchmark exact | 2,368 (37%) | 1,396 (33%) | 980 (41%) |
| An independent evaluatorLabeled Verified | Verified | 295 (5%) | 267 (6%) | 139 (6%) |
| Another lab’s comparison tableLabeled Secondary exact | Secondary exact | 241 (4%) | 237 (6%) | 94 (4%) |
| Carried from an upstream sourceLabeled Reported | Reported | 39 (0%) | 37 (1%) | 1 (0%) |
A rank starts from leaderboards the developer doesn’t run
The results on a model page and the inputs to its rank aren’t the same list. Of the 214 models on the overall board, 185 start from at least one outside leaderboard the developer doesn’t run: a composite benchmark index or a human-preference leaderboard. Benchmark results then move that position by at most 3 points, and they count for more when they come from the benchmark’s own leaderboard than from the developer.
The other 29 have no outside leaderboard yet. We rank them from their benchmark results alone, which can include the developer’s own, and label every one of them Estimated.
A developer’s result can’t make a position Supported
Supported means we have enough independent evidence to stand behind the row, and 79 of the 214 overall positions have it today. Independent evidence is an outside leaderboard, the benchmark’s own leaderboard, or an independent evaluator’s run. A developer’s result is shown and it informs the estimate, but it doesn’t count toward that label. Neither does a result one lab published for another lab’s model.
The rules for both labels, and for how the score is put together, are in how the score is built. What we do about commercial relationships is in the editorial policy.
What stays off the pages
- A result with no public source we can link to isn’t shown. It stays off the page until the table, card or paper it came from is found.
- The outside leaderboards a rank starts from are scored but not republished as rows. Where a category position rests only on them, the model page shows the rank with “Unlisted sources” beside it.
- We don’t run benchmarks ourselves, so nothing here is a first-party measurement.
Found a number that doesn’t match its source? The editorial policy says how to report it, and the model directory is the way into any model’s rows.
Questions
Does BenchLM run its own benchmarks?
No. Each number was published by someone else: the company that built the model, the people who run the benchmark, an independent evaluator, or another lab. Our work is collecting those results, matching each to the exact table or document it came from, and labeling who published it on the row.
How many of the results come from model developers?
As of November 11, 2026, 3,406 of the 6,349 results shown (54%) come from model cards, system cards and launch posts written by the model’s developer. On the pages of the 60 highest-ranked models the share is 49%. These rows carry the label “Provider exact” and link to the document.
Can a developer’s own result move its model up the ranking?
By a bounded amount. 185 of the 214 overall positions start from outside leaderboards, and benchmark results adjust them by 3 points at most. 29 positions rest on benchmark results alone and are marked Estimated. No developer-reported result can make a position Supported.
Where can I check the source of a result?
On the model’s page. Each benchmark row names the document or leaderboard the number came from, links to it, and carries one of five labels: Provider exact, Benchmark exact, Verified, Secondary exact, or Reported. The table on this page says who published the result behind each label.