📊PublicAI Index
PublicAI Index is the LLM benchmark aggregator: one score from every recognised public leaderboard, standardized onto one scale, with launch-post figures marked ✱ and kept out of the headline. It is built from everyone’s benchmarks and none of our own.
Live at publicai.io/model-index.
What it is
Recognised leaderboards (unmarked), fifteen from fourteen publishers: LMArena Text, Artificial Analysis (Intelligence Index, GDPval-AA, AA-Briefcase), Terminal-Bench, ARC-AGI-2, LiveBench, LiveCodeBench, Aider polyglot, BFCL, OSWorld, MMMU-Pro, Kagi, SimpleBench, and Vals.ai’s twenty-two evaluations (legal, finance, tax, medical, cybersecurity, IT operations and more). Read from each publisher’s page; every figure links back to it.
Reports ✱: launch posts and third-party write-ups. Indexed so brand-new models appear before the boards list them, but marked, discounted and never in the Overall ranking.
Catalogs: OpenRouter (callable ids, context, price) and Hugging Face (parameter counts). Never scored.
What it is not
PublicAI runs no benchmarks. We normalize and weight; we do not score.
Weights are fixed and published. They are not user-adjustable.
A snapshot, not a live feed. The date is on every page and in every API response.
Six rules
Standardize, discount uncertainty. Every figure is z-scored across the models its source lists and mapped to 0–100; a published error bar lowers its weight.
Shrink, never impute. A missing figure stays missing; thin evidence is pulled toward 50.
Weight by a published scheme. Each board’s share of the Overall index is fixed and stated with its reason.
Two publishers to rank. Overall needs boards from two independent publishers; a domain ranks on its own board evidence.
Reports ✱ stay out of the headline. They shape domain columns at a tenth of a board’s weight.
Estimate the rest, and say so. A model with no Overall score gets an anchored estimate, shown as ~55✱, never ranked.
See Method for the definitions.
Last updated