Language model rankings / Methodology

How this board works

Marsel measures nothing. Every figure on the board is a measurement published by one of the sources listed below, ingested with its source, its measurement date and — where the source provides one — a link to its own page for that value. The single number Marsel computes is the Marsel score, and its formula is on this page.

The four axes stay separate

The most common way a ranking misleads is by merging things that are not comparable into one number. This board keeps them apart:

Quality: Fixed evaluations and human-preference ratings. This is what the Marsel score is made of, and nothing else is.
Speed: Output throughput and time to first token. Properties of how a model is served — they vary by provider and by load, not only by the model.
Price: Input and output cost per million tokens. Shown against quality as a scatter so the trade-off is yours to make, rather than folded into a score at a weight we chose for you. Prices are carried and displayed AS QUOTED, down to a millionth of a cent: roughly one model in seven costs less than a cent per million tokens, and rounding those to the nearest cent would put a price on this page that the vendor never published.
Usage — popularity, not quality: Token volume routed through a provider. A cheap, fast model wired into one high-traffic application will show more usage than a far more capable model reserved for hard steps. Usage never contributes to the Marsel score. Token counts also come from each provider’s own tokenizer, so they are not strictly comparable between providers.

The Marsel score — formula v1

A weighted mean of normalised quality metrics. The weights below are read directly from the code that computes the score.

Input	Weight	Better when
Human preference (Elo)	30%	higher
Quality index	30%	higher
Coding	20%	higher
Reasoning	20%	higher

1. Normalise. min-max to 0–100 across the ranked cohort. The inputs are on incompatible scales — an index out of 100 beside an Elo around 1300 — so each is mapped to 0–100 against the models currently being ranked, with 100 always meaning best.
2. Weight and average. The weighted mean of whichever inputs a model actually has. weights renormalised over present inputs; below the coverage floor, no score.
3. Refuse to guess. A model with fewer than 2 of the 4 inputs gets no score at all and shows an em-dash. A missing measurement is never scored as zero — that would invent a result nobody produced.

What each metric measures

Quality index (quality): A source's own composite of evaluation results, on its own 0–100 scale. Two sources' indices are not interchangeable.
Human preference (Elo) (quality): A rating derived from blind head-to-head votes by people. It measures which answer humans preferred — not correctness.
Coding (quality): Performance on the source’s coding evaluation set, on its own scale.
Reasoning (quality): Performance on the source’s reasoning evaluation set, on its own scale.
Output speed (speed): Tokens generated per second once output has started. Varies by provider and load.
Latency (TTFT) (speed): Seconds until the first token arrives. Lower is better.
Usage (tokens) (popularity): Token volume routed through a provider. This is POPULARITY, not quality — a cheap model wired into one busy app outranks a more capable one used sparingly. Counts come from each provider’s own tokenizer and are not directly comparable across providers.

Freshness and history

Each source declares how often it refreshes. When one has not delivered within 2× its stated cadence, everything it feeds is marked stale — in the table, under every chart it contributes to, and on the Sources tab with its last-good date. One missed refresh is tolerated; two is reported.
The Trends tab appears only once at least 8 daily snapshots exist. Below that it is absent rather than empty — two points do not make a trend.
Charts plot real observations only. Gaps stay gaps; nothing is interpolated onto a day nobody measured. Price charts step rather than slope, because a price holds until it is changed.
A model missing the plotted metric is excluded from a chart and counted in its “N shown, M excluded” note — never plotted at zero.

Sources

Epoch AI — benchmark and evaluation results · refreshes dailyEpoch AI, "AI Benchmarking Hub". Published online at epoch.ai.
LMArena — human-preference Elo from blind head-to-head votes · refreshes daily
OpenRouter — usage/token volume routed through providers · refreshes dailySource: OpenRouter (openrouter.ai/apps)