The honest leaderboard
for LLM APIs.

MethodFrontier-anchored tiers: each Pareto frontier point of the price/perf plane defines a tier — "best at $X/M in" — and near-frontier active models join their closest anchor by log-price distance. Composite = mean of available subscores: AA mean, llm-stats TrueSkill mean, AA + llm-stats. Before tiering: models are dropped (not shown) if perf < 15 or priced > 6.0x the going rate for their quality (local median price of models within +/-6 perf points); remaining models are grayed when farther than 1.5 sigma from the price/perf Pareto frontier — distance = L2 of the two z-normalized gaps: missing perf at your price, and overpay for your perf.

Cuts"going rate" = median price of models within ±6 quality points. Models > 6.0x it (or quality < 15) are dropped entirely (119 models — old flagships, superseded generations, tiny models); the rest are grayed when more than 1.5σ from the price×quality Pareto frontier (a cheaper AND better model supersedes them). Paid price = OpenRouter traffic-weighted effective price (today, UTC), list price when no traffic.

Data & creditsAll benchmark and pricing data comes from these excellent public resources — go support them: OpenRouter (prices, traffic weighting) · Artificial Analysis (intel/coding/agent indexes) · llm-stats.com (TrueSkill reasoning/coding/agent). Org logos from llm-stats.com. This site is an independent aggregation — all credit to the original data providers.