How to read a ModelCap row

A row brings together an ordering, an estimate and a statement about its evidence. This guide describes Index v4.14 and Score v10, the methods in use since October 2026. Earlier versions read some of these numbers differently, and from 1 to 4 September 2026 they also floored a measured successor above its predecessor. The method changelog on the methodology page records each change, and every model page names the method version that produced its row.

Rank

Rank orders current, canonical language models by their score. Since Index v4.11 the score is the capability point estimate; versions 4.1 to 4.9 ranked on the lower end of the uncertainty interval instead. A tie goes to the row with stronger evidence: measured before inherited before estimated, then higher support. A modeled row is held one tick below a sibling in its own product line whose published results lead it on a strict majority of at least two benchmarks both report, so a line's own ordering is never inverted. Measured and modeled placements both receive ranks, so position alone does not tell you how much independent evidence exists. An available all-versions rank includes eligible superseded releases in the current catalogue; it is not a permanent record of everything that ever ranked.

Score, support, interval

For a measured row, each board contributes its family weight times its source share, scaled by the observation's own confidence after identity, age and sample-size adjustments. An Artificial Analysis result can contribute at most 20 points, an Arena Overall rating at most 60, and the coding, agent and reasoning boards the remaining 20 between them. That is why a launch measured only by Artificial Analysis shows 20% support, however well it scores, until other boards list it. Modeled placements use their own rules: a placement from a model card, for example, gains support with every card benchmark matched against measured peers and never exceeds 40%.

Two things change that width. A row with an Artificial Analysis result is widened on both sides by the held-out error of mapping that index onto the Arena scale, which is why a launch measured only there can show a range such as 73–100. A row without a general board publishes its blend's mean plus or minus 1.28 standard deviations. Endpoints are limited to 0–100, so a displayed interval need not be centred on its score. Read small gaps cautiously and inspect the evidence. These heuristic intervals do not establish a probability that one model is better or demonstrate statistical equivalence.

The evidence label

• Measured. Independent public boards measured the exact catalogue product. It does not mean ModelCap ran the model; nothing on the site is our own benchmark.
• Measured · blended. Measured on specialist boards only, with no general board yet, so the score blends that measurement with modeled estimates that carry more than 40% of the weight.
• Modeled · peer benchmarks. Placed mainly from the publisher's own comparison table against models already on the board, or, when the card names no catalogued peer, from its own reported scores placed against measured models on the same benchmarks. The numbers are quarantined claims; the placement reads them after an optimism haircut and blends them with other estimates.
• Modeled · release claim. A launch with no comparison card yet, placed from where its publisher's own release announcement puts it against one of that publisher's measured models, such as “near GPT-6 Astra”. No number is read from the announcement, and it never anchors another model.
• Modeled · tier sibling. A tier variant such as a Pro or Max model, placed from the measured model it is named after. It never anchors another model.
• Modeled · lineage. Inherited from one verified base model with a disclosed configuration penalty, larger for quantizations. This is inherited support, not a new measurement, and the row always scores below its parent.
• Modeled · succession. Most of the placement comes from an older measured member of the same product family, taken three points below it (or just above it when its card or shared boards show it ahead) and blended with the corpus prior. The result can land above or below that member.
• Modeled · prior. The corpus prior carries most of the blend: the typical score of the publisher's other measured models, or of the whole corpus when there are too few. A prior on its own does not establish a rank, so every such row also has at least one other estimate in its blend.
• Index position pending. No channel can place the model yet, for a written reason: the identity is ambiguous, the card was unreadable, the peers could not be resolved, or the placement was uncalibrated. We show the gap rather than a guess, and every refresh places the model automatically once its card or a public benchmark compares it with ranked models.

Preliminary and Limited evidence

A measured row whose support is under 25% carries a second label. For the first 30 days after release it reads Preliminary, with the number of benchmarks behind the score, because the broad boards usually have not listed the model yet. After that it reads Limited evidence. The label appears on the model page, the releases list and the chart tables. It never moves the rank.

The movement arrow

The movement arrow compares the current rank with an earlier observation. Its tooltip states whether this covers the last 24 hours or the previous daily close. New evidence, catalogue membership and methodology changes can all move ranks; the arrow alone does not identify the cause. The model page's history chart, when available, shows the recorded positions rather than an explanation of each move.

Dates

Snapshot times, source publication dates and catalogue fetch times describe different events. Some sources do not publish row dates, so their recorded observation time is used instead. A source timestamp does not establish when a model was tested, and a catalogue read does not guarantee a price will remain available.

In Score v9 and v10, evidence begins losing freshness support after 60 days, with a 180-day half-life. Observations below 1% combined support leave scoring, and a board with no qualifying recent observation becomes dormant after 90 days. The v4 Index adds a separate stale label and support penalty when even the newest admitted observation is over a year old. Old evidence is not guaranteed to remain in the score.

What a model page adds

The row is a summary. The model page adds source links, admitted evidence, configurations, available timestamps and identity details, plus lineage and quantization information where declared. A repository revision can identify an artifact without proving which revision an evaluator tested. Use the source records and the method version to check the placement; the public dataset provides the underlying published fields for further inspection.