How the LLM Index Score works

Most model leaderboards give you one number and ask you to trust it. This page exists so you do not have to: the inputs are public, the weights are printed below, and every component is shown on every model page. You can recompute any score on this site by hand.

The weights

Coding carries 30% across its two categories because choosing a model for coding work is the most common reason people compare models at all. Reasoning follows at 20% as the broadest proxy for general capability. Mathematics and data analysis sit at 15% each as checks on quantitative reliability, and language and instruction-following round out the remainder.

The formula

categoryScore = mean(raw LiveBench task scores in that category)
LLM Index Score = Σ (categoryScore × categoryWeight)

No renormalisation — LiveBench tasks are already scored 0–100.
A model is scored only if every task in every category is present.

Worked example — OpenAI: GPT-5.5

Sources

• factual OpenRouter models API
• benchmark LiveBench · snapshot 2026_06_25

The LLM Index Score is computed by us from LiveBench's public raw task scores using our own published weights. Third-party scores are attributed to their source and are never presented as ours. Score version v1; index generated 2026-07-21.

Why these weights

Any composite score encodes an opinion about what matters, and the honest thing to do is state the opinion rather than bury it. Ours is that most people comparing models are choosing one to build software with, so the two coding categories together carry the largest share at 30%. Reasoning follows at 20% as the broadest available proxy for general capability — it is the category that most often predicts whether a model handles an unfamiliar problem sensibly.

Mathematics and data analysis sit at 15% each. Both are checks on whether a model is reliable when an answer is verifiably right or wrong, which is exactly where confident-sounding errors do the most damage. Language and instruction-following take 10% each: they matter, but frontier models are now clustered tightly enough there that the category rarely separates them, and weighting it higher would mostly add noise.

You may disagree with all of that, and you have everything you need to act on the disagreement. Every category score is printed on every model page, so you can apply your own weights to the same public numbers and get your own ranking. That is a feature, not a concession — a score you can reweight is a score you can reason about.

What this score does not capture

Benchmarks measure what they measure. This score says nothing about latency, throughput under load, rate limits, regional availability, uptime, safety behaviour, refusal rates, or how a provider handles a traffic spike at 3am. Those decide whether a model is usable in production far more often than a few points of benchmark difference do, and none of them are visible here.

There is also a structural limit worth naming. A benchmark snapshot describes a model as measured on a date. Providers update models behind stable names, so a score can drift from reality without any version number changing. Dating every snapshot is our mitigation, not a cure — treat an older snapshot as evidence about the past, and weight recent measurements more heavily when the gap matters.

Finally, a category mean can hide unevenness. Two models with identical category scores can differ sharply on the individual tasks beneath, which is why the raw task scores are published alongside. If your workload leans on one specific ability, read the task row, not the average.

Versioning and corrections

The score carries a version, currently v1. Any change to the weights increments it and is described on this page. Weights are never adjusted quietly, and never tuned to produce a particular ordering — if a change makes a favoured model look worse, the change still ships. Rankings that move for undisclosed reasons are indistinguishable from rankings that are edited, and the whole value of this index rests on that distinction.

If you find a number here that does not reconcile with the public sources, tell us and we will correct it. A data index that resists correction is just an opinion with tables.

FAQ