One benchmark is one exam.
This is the whole field.

Current rankings

Bar length is the index: a model's position in each category, weighted and averaged. Each band inside a bar is what that category contributed, so you can see whether a lead was won on reasoning or on agents. A category counts for where a model finished against the others who sat the same benchmarks, never for the raw score it posted on them.

Where each model actually ranks

Overall rank is one number and it hides the interesting part. Each cell is this model's rank within that classification, out of every ranked model with enough evidence in it. Ordering is head to head on shared benchmarks only: a model should not gain from sitting easier exams, nor from sitting them against a weaker field, and under the earlier rules it did both. The widest swing on this board is GPT-OSS 120B, 63 percentile points between its best category and its worst. Hover a cell for the record behind it.

Every model with evidence in a category, ranked in it

A coding board needs coding evidence, not all three categories. These boards hold every model with 2 qualifying benchmarks in that one classification, whether or not it has enough spread to carry an overall score. Ordering is head to head on shared benchmarks only. Most of these models are not missing from our reading of the record, they are missing from the world's agent and coding evaluations, which is why so many rank here and nowhere else.

Where the rest of the field is

37 models hold a Ferrox Index. 142 more are measured on at least one category and cannot be scored. A model needs 2 independent benchmarks in EACH of agents, coding, reasoning before it can hold a Ferrox Index. Models below that bar are listed with exactly what they are short of. Almost none of them are unmeasured; they are measured unevenly, and the gap is in the public record, not in our reading of it.

The bottleneck is not close: coding is missing for 125, agents is missing for 125, reasoning is missing for 11 of them. Agent and coding evaluations are run on far fewer models than reasoning evaluations are, and that gap is in the public record rather than in our reading of it. A model here needs one more independently run benchmark in the category shown, and it is comparable.

The score you were shown was chosen

The same model, on the same benchmark, run under a different harness or a different reasoning effort, produces a different score. Whoever publishes gets to pick from that range. Each rail below is the full range one model actually produced, with the value we publish marked inside it and the best configuration marked at the top. We take the median. A best-of rule takes the top mark.

On 32 of the 37 models here, the number moves more than 5 points depending on which run gets quoted. GPT-5.6 Luna spans 65.6 points on DeepSWE across 5 configurations of itself.

Every benchmark, on its own terms

One panel per benchmark, best first, the published value printed on the bar. The leader is in orange and the rest of the field is grey. Under each panel: who ran it, under what licence, and how many ranked models are missing from it. That last line is the one no vendor chart carries.

Behind every bar there is usually more than one number. 321 of these model-and-benchmark cells hold two or more published measurements, and 14 of those hold measurements that disagree while the record says the configuration was the same. The widest gap on the board is 65.6 points, on one model at one runner, moved by nothing but the effort dial. Every claim we can check →

How this is built, and what it refuses

Category ordering is head to head. ordered by Bradley-Terry ratings fitted to every pairwise comparison between two models on the benchmarks they BOTH sat. This orders the category boards, and since 2026-08-29 it orders the Ferrox Index too: a category contributes its position in this fit, weighted by the rubric, never the raw composite it used to contribute. the figure shown is the observed win share: of every head-to-head comparison a model actually played, the share it won. The fitted rating orders the board; it is not published as a score, because under near-perfect transitivity its magnitude reflects the prior rather than a measurement. We rank this way because placement removed the difficulty of a benchmark but not the strength of the field that sat it: the models measured on Terminal-Bench 4.0 averaged overall rank 11.6, while Aider Polyglot averaged 26.8, so sitting an easy room was worth real position. And not on direct results alone because averaging direct results alone compares models over different opponent sets, and 107 of 561 model pairs in the agents category share no benchmark at all. A fitted rating ranks A above C on the strength of both having played B.

every model also plays one drawn game against an average opponent, so a model that won everything has a finite rating and a model with little evidence is pulled gently toward the middle. ratings are normalised so the geometric mean is 1, which pins the scale Bradley-Terry leaves free and makes "an average opponent" mean exactly that. The published figure is the chance of winning one comparison against such an opponent. the comparison graph must be one connected component, which it is in every category on this snapshot, with no isolated model. the fit runs to convergence rather than to a cap, reaching it in 544 to 1833 iterations depending on the category, and the ordering it reaches is identical to one left running 200,000 iterations, in every category. It also barely moves with the strength of the prior: rank correlation against an unregularised fit stays at or above 0.972 from prior 2 to prior 20, four times either side of the value used. Both are asserted by tests rather than assumed. What it does not fix: a benchmark is not a fair coin. Two models separated by a hair and two separated by a mile both count as one win, so this measures who beats whom and not by how much. The margins are on the benchmark panels. placement changes what a category board SHOWS as evidence and never orders anything; the ordering is head to head. Since 2026-08-29 the Ferrox Index is built from those head-to-head positions rather than from the raw category scores, so it is no longer true that the index is untouched by any of this.

How often this board agrees with evidence it never saw: 82.4%. Hide one correlation family of benchmarks. Refit every category board without it, re-derive each category position, re-run the index and re-rank. Then ask the hidden family which of two models is better, and count how often the ordering had already put that one first. Repeat for every family and pool the pairs. An ordering can only be said to have predicted something if it did not already know the answer, so this is the number in the corner and the one to judge the board by. On this snapshot it is measured over 17 folds and 2311 held-out pairs. An arbitrary ordering of the same field scores 51.3% on the same pairs, which is what chance looks like and what this has to be read against. A held-out pair is decided by less evidence than a published pair, often by a single family, so its verdict is noisier and the reachable ceiling is below 100 by an amount this does not estimate. Read it against the arbitrary-ordering baseline, not against 100.

Refitting the board without a family moves it a little and never reorders it wholesale: mean Spearman 0.9992 against the published order across the 17 folds, worst 0.9967. 3 of those folds decided no pair between two ranked models and so validated nothing; they are counted here rather than quietly averaged in as agreement.

And how often it agrees with the evidence it was built from: 95.3%. For every pair of ranked models, count the benchmarks both of them sat and see which won more of them. Pairs that split evenly, or share no benchmark, are set aside as undecided. The agreement rate is the share of the decided pairs this ordering puts the right way round. On this snapshot 619 of 666 pairs are decided that way, and the ordering above puts 29 of them the wrong way round. Those 29 are listed below rather than counted, because a number you cannot check is worth nothing. This figure is measured on the same comparisons the ordering was fitted to, so it is a statement about internal consistency and not about predictive accuracy. It was published here on its own until 2026-08-30, in a position that read as validation, and it should not have been. It counts benchmarks won, never by how much, so a hair and a landslide weigh the same. It cannot reach 100% on evidence that is intransitive, and it has no opinion about which categories matter. It grades orderings against each other; it does not certify one as correct.

The gap between the two is not all overfitting, and reading it that way would be as wrong as the original claim. Graded on one hidden family instead of on everything a pair shares, the published ordering itself scores 83.3%: that drop is the price of thinner evidence per pair, and it is paid whether or not the ordering saw the fold. Taking the fold out of the fit costs a further 0.9 points, on identical pairs with nothing else varied. That last number is the overfitting, and it is the one worth arguing about.

Showing the 10 widest of 29. An ordering cannot satisfy every pair when the evidence itself is intransitive, so this number will never reach 100 and a board claiming it did would be telling you something else.

A grade is a position, never an absolute. The field on this snapshot runs from 99.5 down to 23.1, so grading against a 0 to 100 scale would put the best model in the world in the middle of the alphabet. a grade boundary never splits a pair the evidence cannot separate; the lower row is lifted, capped at one band, and the lift is named on the row.