LLM List methodology

LLM List combines model and provider facts from models.dev, OpenRouter, and Arena; benchmark measurements from Artificial Analysis; and open-weight metadata from Hugging Face. It refreshes every source every six hours and retains each last-good response.

Model identity

Provider catalogues publish one record per endpoint. We fold equivalent names and verified delivery aliases—such as provider, region or accelerated routes—into one canonical model and retain every endpoint as an offering variant. OpenRouter's own API replaces its models.dev offering with the fresher direct list. Arena's public catalogue is deduplicated by public model ID; anonymous stealth routes are excluded. Product tiers and model versions remain separate. Capabilities are the union across offerings; limits use the most commonly published value so one outlier does not become the headline.

Pricing

The headline input/output pair always comes from one real paid endpoint, selected by combined price. A free endpoint sets a separate free-tier flag. Prices are dollars per million tokens and should be verified with the provider.

Benchmarks

Models are joined to Artificial Analysis by normalized name. Every reasoning configuration is retained. The list uses the best intelligence run and the throughput from that same run. Individual benchmark fractions are normalized to a 0–100 scale.

Price history

Every refresh observes each model's headline price. A new point is stored only when the price changes, producing an exact step history rather than a sampled line. A new deployment begins with one observation and draws a chart after the first change.

Local hardware estimates

Open-weight models are conservatively matched to a Hugging Face repository. Working-memory estimates adapt the MIT-licensed whichllm formulas at a pinned revision and include weights, 4K KV cache, activation/scratch memory and framework overhead. They are planning estimates, not minimum requirements. Open weights does not mean open source; consult the model card and its displayed license.

Limitations and sources

Supplemental provider facts, scores and hardware facts do not affect catalogue availability. Upstream facts can be incomplete or delayed. This site attributes source data and presents its matching and folding rules openly; verify prices, limits and capabilities before relying on them.