The AI Model Index

A ranking of 145 large language models built by cross-referencing three independent sources: live market pricing from OpenRouter, benchmark indices from Artificial Analysis, and human preference votes from LMArena. We aggregate those measurements and compute our own score across five weighted axes. The raw measurements belong to their sources; the scoring and the ranking are ours.

By measured capability, not by price. A model that costs ten times more but answers harder questions ranks higher by default. Three alternative orderings are one click away — price/performance, speed, and the weighted composite score — because "best" means something different depending on whether you are paying the bill or waiting for the response.

Two things, weighted equally when both exist. Machine benchmarks contribute the general intelligence index (50%), the coding index (25%) and the agentic index (25%), each normalised against its own ceiling. Human preference contributes the LMArena Elo rating, mapped onto the same 0-100 scale. A model measured by only one of the two is scored on that one alone.

A model carrying both a machine benchmark score and a human preference rating. 14 of the 145 models in the index qualify. It matters because the two methods disagree more often than you would expect: benchmarks measure what is testable, the arena measures what people actually prefer after using the model.

They measure different things. Benchmarks test verifiable problems with known answers — maths, code that compiles, factual recall. Human votes reward tone, instruction-following, formatting and refusal behaviour, which no benchmark captures. A model can top the benchmarks and lose blind head-to-head votes, and the reverse happens just as often.

Because nearly every current model scored in the 90s on it, so it separated nothing. Worse, averaging it in rewarded models measured on maths and almost nothing else — one model with no coding or agentic score reached a higher intelligence score than models measured on everything. Artificial Analysis has since stopped publishing that index on the endpoint we use, which changes nothing for us: it was already out of the score.

Measured capability divided by the blended price per million tokens, weighted one part input to three parts output, which reflects typical real-world usage. The result is placed on a logarithmic scale because prices span four orders of magnitude. Free or unpriced models receive a neutral 50 rather than a perfect score, since they are not comparable.

Yes. They come from OpenRouter, which publishes live per-token rates charged by the providers actually serving each model, in US dollars per million tokens. Input and output are shown separately because the gap between them is often large and matters more than the headline figure.

Every 48 hours. Each collection re-fetches all three sources, recalculates every score, and writes a historical snapshot so price changes and ranking movements stay traceable. Models that disappear from the sources are marked inactive rather than deleted, so their history survives.

A model whose weights can be downloaded and self-hosted, as opposed to one reachable only through a vendor API. 85 of the 145 models qualify. Open weights add points on the ecosystem axis because they remove vendor lock-in, but a closed model served by many providers is not penalised to zero.

So the numbers stay comparable over time. If scores were normalised against the current field, every new model release would silently move every existing score, and a snapshot from six months ago would mean nothing. On an absolute scale, a model scoring 78 today means the same as a model scoring 78 last year.

16 of the 145 models link to a full review on this site, written after actually using the model. Those links appear directly on each model card and in the comparison table. The rest are scored from published measurements only, which is stated rather than hidden.